Post-Modern Machine Learning: The Deep Revolution

The mysteries of human intelligence are among the final frontiers of science. Despite our pride in what science has achieved across the past century, we have stalled when it comes to understanding intelligence or emulating it. The best we have done so far is through machine learning — harnessing the computational power of computers to begin to mimic the mind, attempting to answer the essential question:

How do we get machines to Know what we Know?

In modern machine learning, the answer is algorithmic.

In post-modern machine learning, the answer is manifestation.

The algorithms of modern machine learning are cause and effect, rules to follow, producing only what the programmer can imagine. But post-modern machine learning has blown past explicit algorithms to embrace deep networks. Deep networks today are defined by neural networks with thousands, or tens of thousands, or even hundreds of thousands, of neurons arrayed in multiple layers of dense connections. The interactivity of so many crossing streams of information defies direct deconstruction of what the networks are doing — they are post-modern. Their outputs manifest themselves, self-assembling into simplified structures and patterns and dependencies that are otherwise buried unseen in complicated data.

Fig. 1 A representation of a deep network with three fully-connected hidden layers. Deep networks are typically three or more layers deep, but each layer can have thousands of neurons. (Figure from the TowardsDataScience blog.)

Deep learning emerged as recently as 2006 and has opened wide new avenues of artificial intelligence that move beyond human capabilities for some tasks.  But deep learning also has pitfalls, some of which are inherited from the legacy approaches of traditional machine learning, and some of which are inherent in the massively high-dimensional spaces in which deep learning takes place.  Nonetheless, deep learning has revolutionized many aspects of science, and there is reason for optimism that the revolution will continue. Fifty years from now, looking back, we may recognize this as the fifth derivative of the industrial revolution (Phase I: Steam. Phase II: Electricity. Phase III: Automation. Phase IV: Information. Phase V: Intelligence).

From Multivariate Analysis to Deep Learning

Conventional machine learning, as we know it today, has had many names.  It began with Multivariate Analysis of mathematical population dynamics around the turn of the last century, pioneered by Francis Galton (1874), Karl Pearson (1901), Charles Spearman (1904) and Ronald Fisher (1922) among others.

The first on-line computers during World War II were developed to quickly calculate the trajectories of enemy aircraft for gunnery control, introducing the idea of feedback control of machines. This was named Cybernetics by Norbert Wiener, who had participated in the development of automated control of antiaircraft guns.

Table I. Evolution of Names for Machine Learning

A decade later, during the Cold War, it became necessary to find hidden objects in large numbers of photographs.  The embryonic electronic digital computers of the day were far too slow with far too little memory to do the task, so the Navy contracted with the Cornell Aeronautical Laboratory in Cheektowaga, New York, a suburb of Buffalo, to create an analog computer capable of real-time image analysis.  This led to the invention of the Perceptron by Frank Rosenblatt as the first neural network-inspired computer [1], building on ideas of neural logic developed by Warren McColloch and Walter Pitts. 

Fig. 2 Frank Rosenblatt working on the Perceptron. (From the Cornell Chronicle)
Fig. 3 Rosenblatt’s conceptual design of the connectionism of the perceptron (1958).

Several decades passed with fits and starts as neural networks remained too simple to accomplish anything profound.  Then in 1986, David Rumelhart and Ronald Williams at UC San Diego with Geoff Hinton at Carnegie-Mellon discovered a way to train multiple layers of neurons, in a process called error back propagation [2].  This publication opened the floodgates of Connectionism — also known as Parallel Distributed Processing.  The late 80’s and much of the 90’s saw an expansion of interest in neural networks, until the increasing size of the networks ran into limits caused by the processing speed and capacity of conventional computers towards the end of the decade.  During this time it had become clear that the most interesting computations required many layers of many neurons, and the number of neurons expanded into the thousands, but it was difficult to train such systems that had tens of thousands of adjustable parameters, and research in neural networks once again went into a lull.

The beginnings of deep learning started with two breakthroughs.  The first was by Yann Lecun at Bell Labs in 1998 who developed, with Leon Bottou, Yoshua Bengio and Patrick Haffner, a Convolutional Neural Network that had seven layers of neurons that classified hand-written digits [3]. The second was from Geoff Hinton in 2006, by then at the University of Toronto, who discovered a fast learning algorithm for training deep layers of neurons [4].  By the mid 2010’s, research on neural networks was hotter than ever, propelled in part by several very public successes, such as Deep Mind’s machine that beat the best player in the world at Go in 2017, self-driving cars, personal assistants like Siri and Alexa, and YouTube recommendations.

The Challenges of Deep Learning

Deep learning today is characterized by neural network architectures composed of many layers of many neurons.  The nature of deep learning brings with it two main challenges:  1) efficient training of the neural weights, and 2) generalization of trained networks to perform accurately on previously unseen data inputs.

Solutions to the first challenge, efficient training, are what allowed the deep revolution in the first place—the result of a combination of increasing computer power with improvements in numerical optimization. This included faster personal computers that allowed nonspecialists to work with deep network programming environments like Matlab’s Deep Learning toolbox and Python’s TensorFlow.

Solutions to the second challenge, generalization, rely heavily on a process known as “regularization”. The term “regularization” has a slippery definition, an obscure history, and an awkward sound to it. Regularization is the noun form of the verb “to regularize” or “to make regular”. Originally, regularization was used to keep certain inverse algorithms from blowing up, like inverse convolutions, also known as deconvolution. Direct convolution is a simple mathematical procedure that “blurs” ideal data based on the natural response of a measuring system. However, if one has experimental data, one might want to deconvolve the system response from the data to recover the ideal data. But this procedure is numerically unstable and can “blow up”, often because of the divide-by-zero problem. Regularization was a simple technique that kept denominators from going to zero.

Regularization became a common method for inverse problems that are notorious to solve because of the many-to-one mapping that can occur in measurement systems. There can be a number of possible causes that produce a single response. Regularization was a way of winnowing out “unphysical” solutions so that the physical inverse solution remained.

During the same time, regularization became a common tool used by quantum field theorists to prevent certain calculated quantities from diverging to infinity. The solution was again to keep denominators from going to zero by setting physical cut-off lengths on the calculations. These cut-offs were initially ad hoc, but the development of renormalization group theory by Kenneth Wilson at Cornell (Nobel Prize in 1982) provided a systematic approach to solving the infinities of quantum field theory.

With the advent of neural networks, having hundreds to thousands to millions of adjustable parameters, regularization became the catch-all term for fighting the problem of over-fitting. Over-fitting occurs when there are so many adjustable parameters that any training data can be fit, and the neural network becomes just a look-up table. Look-up tables are the ultimate hash code, but they have no predictive capability. If a slightly different set of data are fed into the network, the output can be anything. In over-fitting, there is no generalization, the network simply learns the idiosyncrasies of the training data without “learning” the deeper trends or patterns that would allow it to generalize to handle different inputs.

Over the past decades a wide collection of techniques have been developed to reduce over-fitting of neural networks. These techniques include early stopping, k-fold holdout, drop-out, L1 and L2 weight-constraint regularization, as well as physics-based constraints. The goal of all of these techniques is to keep neural nets from creating look-up tables and instead learning the deep codependencies that exist within complicated data.

Table II. Regularization Techniques in Machine Learning

By judicious application of these techniques, combined with appropriate choices of network design, amazingly complex problems can be solved by deep networks and they can generalized too (to some degree). As the field moves forward, we may expect additional regularization tricks to improve generalization, and design principles will emerge so that the networks no longer need to be constructed by trial and error.

The Potential of Deep Learning

In conventional machine learning, one of the most critical first steps performed on a dataset has been feature extraction. This step is complicated and difficult, especially when the signal is buried either in noise or in confounding factors (also known as distractors). The analysis is often highly sensitive to the choice of features, and the selected features may not even be the right ones, leading to bad generalization. In deep learning, feature extraction disappears into the net itself. Optimizing the neural weights subject to appropriate constraints forces the network to find where the relevant information lies and what to do with it.

The key to finding the right information was not just having many neurons, but having many layers, which is where the potential of deep learning emerges. It is as if each successive layer is learning a more abstract or more generalized form of the information than the last. This hierarchical layering is most evident in the construction of convolutional deep networks, where the layers are receiving a telescoping succession of information fields from lower levels. Geoff Hinton‘s Deep Belief Network, which launched the deep revolution in 2006, worked explicitly with this hierarchy in mind through direct design of the network architecture. Since then, network architecture has become more generalized, with less up-front design while relying on the increasingly sophisticated optimization techniques of training to set the neural weights. For instance, a simplified instance of a deep network is shown in Fig. 4 with three hidden layers of neurons.

Fig. 4 General structure of a deep network with three hidden layers. Layers will typically have hundreds or thousands of neurons. Each gray line represents a weight value, and each circle is a neural activation function.

The mathematical structure of a deep network is surprisingly simple. The equations for the network in Fig. 4, that convert an input xa to an output ye, are

These equations use index notation to denote vectors (single superscript) and matrices (double indexes). The repeated index (one up and one down) denotes an implicit “Einstein” summation. The function φ(.) is known as the activation function, which is nonlinear. One of the simplest activation functions to use and analyze, and the current favorite, is known as the ReLU (for rectifying linear unit). Note that these equations represent a simple neural cascade, as the output of one layer becomes the input for the next.

The training of all the matrix elements assumes a surprisingly simply optimization function, known as an objective function or a loss function, that can look like

where the first term is the mean squared error of the network output ye relative to the desired output y0 for the training set, and the second term, known as a regularization term (see section above) is a quadratic form that keeps the weights from blowing up. This loss function is minimized over the set of adjustable matrix weights.

The network in Fig. 4 is just a toy, with only 5 inputs and 5 outputs and only 23 neurons. But it has 30+36+36+30+23 = 155 adjustable weights. If this seems like overkill, but it is nothing compared to neural networks with thousands of neurons per layer and tens of layers. That massive overkill is exactly the power of deep learning — as well as its pitfall.

The Pitfalls of Deep Learning

Despite the impressive advances in deep learning, serious pitfalls remain for practitioners. One of the most challenging problems in deep learning is the saddle-point problem. A saddle-point in an objective function is like a mountain pass through the mountains: at the top of the pass it slopes downward in two opposite directions into the valleys but slopes upward in the two orthogonal directions to mountain peaks. A saddle point is an unstable equilibrium where a slight push this way or that can lead the traveller to two very different valleys separated by high mountain ridges. In our familiar three-dimensional space, saddle points are relatively rare and landscapes are dominated by valleys and mountaintops. But this intuition about landscapes fails in high dimensions.

Landscapes in high dimensions are dominated by neutral ridges that span the domain of the landscape. This key understanding about high-dimensional space actually came from the theory of evolutionary dynamics for the evolution of species. In the early days of evolutionary dynamics, there was a serious challenge to understand how genetic mutations could allow such diverse speciation to occur. If the fitness of a species were viewed as a fitness landscape, and if a highly-adapted species were viewed as a mountain peak in this landscape, then genetic mutations would need to drive the population state point into “valleys of death” that would need to be crossed to arrive at a neighboring fitness peak. It would seem that genetic mutations would likely kill off the species in the valleys before they could rise to the next equilibrium.

However, the geometry of high dimensions does not follow this simple low-dimensional intuition. As more dimensions come available, landscapes have more and more ridges of relatively constant height that span the full space (See my recent blog on random walks in 10-dimensions and my short YouTube video). For a species to move from one fitness peak to another fitness peak in a fitness landscape (in ultra-high-dimensional mutation space), all that is needed is for a genetic mutation to step the species off of the fitness peak onto a neutral ridge where many mutations can keep the species on that ridge as it moves ever farther away from the last fitness peak. Eventually, the neutral ridge brings the species near a new fitness peak where it can climb to the top, creating a new stable species. The point is that most genetic mutations are neutral — they do not impact the survivability of an individual. This is known as the neutral network theory of evolution proposed by Motoo Kimura (1924 – 1994) [5]. As these mutation accumulate, the offspring can get genetically far from the progenitor. And when a new fitness peak comes near, many of the previously neutral mutations can come together and become a positive contribution to fitness as the species climbs the new fitness peak.

The neutral network of genetic mutation was a paradigm shift in the field of evolutionary dynamics, and it also taught everyone how different random walks in high-dimensional spaces are from random walks in 3D. But although neutral networks solved the evolution problem, they become a two-edged sword in machine learning. On the positive side, fitness peaks are just like the minima of objective functions, and the ability for partial solutions to perform random walks along neutral ridges in the objective-function space allows optimal solutions to be found across a broad range of the configuration space of the neural weights. However, on the negative side, ridges are loci of unstable equilibrium. Hence there are always multiple directions that a solution state can go to minimize the objective function. Each successive run of a deep-network neural weight optimizer can find equivalent optimal solutions — but they each can be radically different. There is no hope of averaging the weights of an ensemble of networks to arrive at an “average” deep network. The averaging would simply drive all weights to zero. Certainly, the predictions of an ensemble of equivalently trained networks can be averaged—but this does not illuminate what is happening “under the hood” of the machine, which is where our own “understanding” of what the network is doing would come from.

Post-Modern Machine Learning

Post-modernism is admittedly kind of a joke — it works so hard to pull down every aspect of human endeavor that it falls victim to its own methods. The convoluted arguments made by its proponents sound like ultimate slacker talk — circuitous logic circling itself in an endless loop of denial.

But post-modernism does have its merits. It surfs on the moving crest of what passes as modernism, as modernism passes onward to its next phase. The philosophy of post-modernism moves beyond rationality in favor of a subjectivism in which cause and effect are blurred.  For instance, in post-modern semiotic theory, a text or a picture is no longer an objective element of reality, but fragments into multiple semiotic versions, each one different for each different reader or watcher — a spectrum of collaborative efforts between each consumer and the artist. The reader brings with them a unique set of life experiences that interact with the text to create an entirely new experience in each reader’s mind.

Deep learning is post-modern in the sense that deterministic algorithms have disappeared. Instead of a traceable path of sequential operations, neural nets scramble information into massively-parallel strings of partial information that cross and interact nonlinearly with other massively-parallel strings. It is difficult to impossible to trace any definable part of the information from input to output. The output simply manifests some aspect of the data that was hidden from human view.

But the Age of Artificial Intelligence is not here yet. The vast multiplicity of saddle ridges in high dimensions is one of the drivers for one of the biggest pitfalls of deep learning — the need for massive amounts of training data. Because there are so many adjustable parameters in a neural network, and hence so many dimensions, a tremendous amount of training data is required to train a network to convergence. This aspect of deep learning stands in strong contrast to human children who can be shown a single picture of a bicycle next to a tricycle, and then they can achieve almost perfect classification accuracy when shown any number of photographs of different bicycles and tricycles. Humans can generalize with an amazingly small amount of data, while deep networks often need thousands of examples. This example alone points to the marked difference between human intelligence and the current state of deep learning. There is still a long road ahead.


[1] F. Rosenblatt, “THE PERCEPTRON – A PROBABILISTIC MODEL FOR INFORMATION-STORAGE AND ORGANIZATION IN THE BRAIN,” Psychological Review, vol. 65, no. 6, pp. 386-408, (1958)

[2] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “LEARNING REPRESENTATIONS BY BACK-PROPAGATING ERRORS,” Nature, vol. 323, no. 6088, pp. 533-536, Oct (1986)

[3] LeCun, Yann; Léon Bottou; Yoshua Bengio; Patrick Haffner (1998). “Gradient-based learning applied to document recognition”. Proceedings of the IEEE. 86 (11): 2278–2324.

[4] G. E. Hinton, S. Osindero, and Y. W. Teh, “A fast learning algorithm for deep belief nets,” Neural Computation, vol. 18, no. 7, pp. 1527-1554, Jul (2006)

[5] M. Kimura, The Neutral Theory of Molecular Evolution. Cambridge University Press, 1968.

Twenty Years at Light Speed: Photonic Computing

In the epilog of my book Mind at Light Speed: A New Kind of Intelligence (Free Press, 2001), I speculated about a future computer in which sheets of light interact with others to form new meanings and logical cascades as light makes decisions in a form of all-optical intelligence.

Twenty years later, that optical computer seems vaguely quaint, not because new technology has passed it by, like looking at the naïve musings of Jules Verne from our modern vantage point, but because the optical computer seems almost as far away now as it did back in 2001.

At the the turn of the Millennium we were seeing tremendous advances in data rates on fiber optics (see my previous Blog) as well as the development of new types of nonlinear optical devices and switches that served the role of rudimentary logic switches.  At that time, it was not unreasonable to believe that the pace of progress would remain undiminished, and that by 2020 we would have all-optical computers and signal processors in which the same optical data on the communication fibers would be involved in the logic that told the data what to do and where to go—all without the wasteful and slow conversion to electronics and back again into photons—the infamous OEO conversion.

However, the rate of increase of the transmission bandwidth on fiber optic cables slowed not long after the publication of my book, and nonlinear optics today still needs high intensities to be efficient, which remains a challenge for significant (commercial) use of all-optical logic.

That said, it’s dangerous to ever say never, and research into all-optical computing and data processing is still going strong (See Fig. 1).  It’s not the dream that was wrong, it was the time-scale that was wrong, just like fiber-to-the-home.  Back in 2001, fiber-to-the-home was viewed as a pipe-dream by serious technology scouts.  It took twenty years, but now that vision is coming true in urban settings.  Back in 2001, all-optical computing seemed about 20 years away, but now it still looks 20 years out.  Maybe this time the prediction is right.  Recent advances in all-optical processing give some hope for it.  Here are some of those advances.

Fig. 1 Number of papers published by year with the phrase in the title: “All-Optical” or “Photonic or Optical and Neur*” according to Web of Science search. The term “All-optical” saturated around 2005. Papers written around optical neural networks was low to 2015 but now is experiencing a strong surge. The sociology of title choices, and how favorite buzz words shift over time, can obscure underlying causes and trends, but overall there is current strong interest in all-optical systems.

The “What” and “Why” of All-Optical Processing

One of the great dreams of photonics is the use of light beams to perform optical logic in optical processors just as electronic currents perform electronic logic in transistors and integrated circuits. 

Our information age, starting with the telegraph in the mid-1800’s, has been built upon electronics because the charge of the electron makes it a natural decision maker.  Two charges attract or repel by Coulomb’s Law, exerting forces upon each other.  Although we don’t think of currents acting in quite that way, the foundation of electronic logic remains electrical interactions. 

But with these interactions also come constraints—constraining currents to be contained within wires, waiting for charging times that slow down decisions, managing electrical resistance and dissipation that generate heat (computer processing farms in some places today need to be cooled by glacier meltwater).  Electronic computing is hardly a green technology.

Therefore, the advantages of optical logic are clear: broadcasting information without the need for expensive copper wires, little dissipation or heat, low latency (signals propagate at the speed of light).  Furthermore, information on the internet is already in the optical domain, so why not keep it in the optical domain and have optical information packets making the decisions?  All the routing and switching decisions about where optical information packets should go could be done by the optical packets themselves inside optical computers.

But there is a problem.  Photons in free space don’t interact—they pass through each other unaffected.  This is the opposite of what is needed for logic and decision making.  The challenge of optical logic is then to find a way to get photons to interact.

Think of the scene in Star Wars: The New Hope when Obiwan Kenobi and Darth Vader battle to the death in a light saber duel—beams of light crashing against each other and repelling each other with equal and opposite forces.  This is the photonic engineer’s dream!  Light controlling light.  But this cannot happen in free space. On the other hand, light beams can control other light beams inside nonlinear crystals where one light beam changes the optical properties of the crystal, hence changing how another light beam travels through it.  These are nonlinear optical crystals.

Nonlinear Optics

Virtually all optical control designs, for any kind of optical logic or switch, require one light beam to affect the properties of another, and that requires an intervening medium that has nonlinear optical properties.  The physics of nonlinear optics is actually simple: one light beam changes the electronic structure of a material which affects the propagation of another (or even the same) beam.  The key parameter is the nonlinear coefficient that determines how intense the control beam needs to be to produce a significant modulation of the other beam.  This is where the challenge is.  Most materials have very small nonlinear coefficients, and the intensity of the control beam usually must be very high. 

Fig. 2 Nonlinear optics: Light controlling light. Light does not interact in free space, but inside a nonlinear crystal, polarizability can create an effect interaction that can be surprisingly strong. Two-wave mixing (exchange of energy between laser beams) is shown in the upper pane. Optical associative holographic memory (four-wave mixing) is an example of light controlling light. The hologram is written when exposed by both “Light” and “Guang/Hikari”. When the recorded hologram is presented later only with “Guang/Hikari” it immediately translates it to “Light”, and vice versa.

Therefore, to create low-power all-optical logic gates and switches there are four main design principles: 1) increase the nonlinear susceptibility by engineering the material, 2) increase the interaction length between the two beams, 3) concentrate light into small volumes, and 4) introduce feedback to boost the internal light intensities.  Let’s take these points one at a time.

Nonlinear susceptibility: The key to getting stronger interaction of light with light is in the ease with which a control beam of light can distort the crystal so that the optical conditions change for a signal beam. This is called the nonlinear susceptibility . When working with “conventional” crystals like semiconductors (e.g. CdZnSe) or rare-Earths (e.g. LiNbO3), there is only so much engineering that is possible to try to tweak the nonlinear susceptibilities. However, artificially engineered materials can offer significant increases in nonlinear susceptibilities, these include plasmonic materials, metamaterials, organic semiconductors, photonic crystals. An increasingly important class of nonlinear optical devices are semiconductor optical amplifiers (SOA).

Interaction length: The interaction strength between two light waves is a product of the nonlinear polarization and the length over which the waves interact. Interaction lengths can be made relatively long in waveguides but can be made orders of magnitude longer in fibers. Therefore, nonlinear effects in fiber optics are a promising avenue for achieving optical logic.

Intensity Concentration:  Nonlinear polarization is the product of the nonlinear susceptibility with the field amplitude of the waves. Therefore, focusing light down to small cross sections produces high power, as in the core of a fiber optic, again showing advantages of fibers for optical logic implementations.

Feedback: Feedback, as in a standing-wave cavity, increases the intensity as well as the effective interaction length by folding the light wave continually back on itself. Both of these effects boost the nonlinear interaction, but then there is an additional benefit: interferometry. Cavities, like a Fabry-Perot, are interferometers in which a slight change in the round-trip phase can produce large changes in output light intensity. This is an optical analog to a transistor in which a small control current acts as a gate for an exponential signal current. The feedback in the cavity of a semiconductor optical amplifier (SOA), with high internal intensities and long effective interaction lengths and an active medium with strong nonlinearity make these elements attractive for optical logic gates. Similarly, integrated ring resonators have the advantage of interferometric control for light switching. Many current optical switches and logic gates are based on SOAs and integrated ring resonators.

All-Optical Regeneration

The vision of the all-optical internet, where the logic operations that direct information to different locations is all performed by optical logic without ever converting into the electrical domain, is facing a barrier that is as challenging to overcome today as it was back in 2001: all-optical regeneration. All-optical regeneration has been and remains the Achilles Heal of the all-optical internet.

Signal regeneration is currently performed through OEO conversion: Optical-to-Electronic-to-Optical. In OEO conversion, a distorted signal (distortion is caused by attenuation and dispersion and noise as signals travel down fiber optics) is received by a photodetector, is interpreted as ones and zeros that drive laser light sources that launch the optical pulses down the next stretch of fiber. The new pulses are virtually perfect, but they again degrade as they travel, until they are regenerated, and so on. The added advantage of the electrical layer is that the electronic signals can be used to drive conventional electronic logic for switching.

In all-optical regeneration, on the other hand, the optical pulses need to be reamplified, reshaped and retimed––known as 3R regeneration––all by sending the signal pulses through nonlinear amplifiers and mixers, which may include short stretches of highly nonlinear fiber (HNLF) or semiconductor optical amplifiers (SOA). There have been demonstrations of 2R all-optical regeneration (reamplifying and reshaping but not retiming) at lower data rates, but getting all 3Rs at the high data rates (40 Gb/s) in the next generation telecom systems remains elusive.

Nonetheless, there is an active academic literature that is pushing the envelope on optical logical devices and regenerators [1]. Many of the systems focus on SOA’s, HNLF’s and Interferometers. Numerical modeling of these kinds of devices is currently ahead of bench-top demonstrations, primarily because of the difficulty of fabrication and device lifetime. But the numerical models point to performance that would be competitive with OEO. If this OOO conversion (Optical-to-Optical-to-Optical) is scalable (can handle increasing bit rates and increasing numbers of channels), then the current data crunch that is facing the telecom trunk lines (see my previous Blog) may be a strong driver to implement such all-optical solutions.

It is important to keep in mind that legacy technology is not static but also continues to improve. As all-optical logic and switching and regeneration make progress, OEO conversion gets incrementally faster, creating a moving target. Therefore, we will need to wait another 20 years to see whether OEO is overtaken and replaced by all-optical.

Fig. 3 Optical-Electronic-Optical regeneration and switching compared to all-optical control. The optical control is performed using SOA’s, interferometers and nonlinear fibers.

Photonic Neural Networks

The most exciting area of optical logic today is in analog optical computing––specifically optical neural networks and photonic neuromorphic computing [2, 3]. A neural network is a highly-connected network of nodes and links in which information is distributed across the network in much the same way that information is distributed and processed in the brain. Neural networks can take several forms––from digital neural networks that are implemented with software on conventional digital computers, to analog neural networks implemented in specialized hardware, sometimes also called neuromorphic computing systems.

Optics and photonics are well suited to the analog form of neural network because of the superior ability of light to form free-space interconnects (links) among a high number of optical modes (nodes). This essential advantage of light for photonic neural networks was first demonstrated in the mid-1980’s using recurrent neural network architectures implemented in photorefractive (nonlinear optical) crystals (see Fig. 1 for a publication timeline). But this initial period of proof-of-principle was followed by a lag of about 2 decades due to a mismatch between driver applications (like high-speed logic on an all-optical internet) and the ability to configure the highly complex interconnects needed to perform the complex computations.

Fig. 4 Optical vector-matrix multiplication. An LED array is the input vector, focused by a lens onto the spatial light modulator that is the 2D matrix. The transmitted light is refocussed by the lens onto a photodiode array with is the output vector. Free-space propagation and multiplication is a key advantage to optical implementation of computing.

The rapid rise of deep machine learning over the past 5 years has removed this bottleneck, and there has subsequently been a major increase in optical implementations of neural networks. In particular, it is now possible to use conventional deep machine learning to design the interconnects of analog optical neural networks for fixed tasks such as image recognition [4]. At first look, this seems like a non-starter, because one might ask why not use the conventional trained deep network to do the recognition itself rather than using it to create a special-purpose optical recognition system. The answer lies primarily in the metrics of latency (speed) and energy cost.

In neural computing, approximately 90% of the time and energy go into matrix multiplication operations. Deep learning algorithms driving conventional digital computers need to do the multiplications at the sequential clock rate of the computer using nested loops. Optics, on the other had, is ideally suited to perform matrix multiplications in a fully parallel manner (see Fig. 4). In addition, a hardware implementation using optics operates literally at the speed of light. The latency is limited only by the time of flight through the optical system. If the optical train is 1 meter, then the time for the complete computation is only a few nanoseconds at almost no energy dissipation. Combining the natural parallelism of light with the speed has led to unprecedented computational rates. For instance, recent implementations of photonic neural networks have demonstrated over 10 Trillion operations per second (TOPS) [5].

It is important to keep in mind that although many of these photonic neural networks are characterized as all-optical, they are generally not reconfigurable, meaning that they are not adaptive to changing or evolving training sets or changing input information. Most adaptive systems use OEO conversion with electronically-addressed spatial light modulators (SLM) that are driven by digital logic. Another technology gaining recent traction is neuromorphic photonics in which neural processing is implemented on photonic integrated circuits (PICS) with OEO conversion. The integration of large numbers of light emitting sources on PICs is now routine, relieving the OEO bottleneck as electronics and photonics merge in silicon photonics.

Farther afield are all-optical systems that are adaptive through the use of optically-addressed spatial light modulators or nonlinear materials. In fact, these types of adaptive all-optical neural networks were among the first demonstrated in the late 1980’s. More recently, advanced adaptive optical materials, as well as fiber delay lines for a type of recurrent neural network known as reservoir computing, have been used to implement faster and more efficient optical nonlinearities needed for adaptive updates of neural weights. But there are still years to go before light is adaptively controlling light entirely in the optical domain at the speeds and with the flexibility needed for real-world applications like photonic packet switching in telecom fiber-optic routers.

In stark contrast to the status of classical all-optical computing, photonic quantum computing is on the cusp of revolutionizing the field of quantum information science. The recent demonstration from the Canadian company Xanadu of a programmable photonic quantum computer that operates at room temperature may be the harbinger of what is to come in the third generation Machines of Light: Quantum Optical Computers, which is the topic of my next blog.

Further Reading

[1] V. Sasikala and K. Chitra, “All optical switching and associated technologies: a review,” Journal of Optics-India, vol. 47, no. 3, pp. 307-317, Sep (2018)

[2] C. Huang et a., “Prospects and applications of photonic neural networks,” Advances in Physics-X, vol. 7, no. 1, Jan (2022), Art no. 1981155

[3] G. Wetzstein, A. Ozcan, S. Gigan, S. H. Fan, D. Englund, M. Soljacic, C. Denz, D. A. B. Miller, and D. Psaltis, “Inference in artificial intelligence with deep optics and photonics,” Nature, vol. 588, no. 7836, pp. 39-47, Dec (2020)

[4] X. Lin, Y. Rivenson, N. T. Yardimei, M. Veli, Y. Luo, M. Jarrahi, and A. Ozcan, “All-optical machine learning using diffractive deep neural networks,” Science, vol. 361, no. 6406, pp. 1004-+, Sep (2018)

[5] X. Y. Xu, M. X. Tan, B. Corcoran, J. Y. Wu, A. Boes, T. G. Nguyen, S. T. Chu, B. E. Little, D. G. Hicks, R. Morandotti, A. Mitchell, and D. J. Moss, “11 TOPS photonic convolutional accelerator for optical neural networks,” Nature, vol. 589, no. 7840, pp. 44-+, Jan (2021)