Neural Networks on Photonic Chips

More on this

Neural networks on photonic chips: Harnessing light for ultra-fast and low-power artificial intelligence

by Polytechnic University of Milan

Neural networks on photonic chips. Credit: Politecnico di Milano

Neural networks are distributed computing structures inspired by the structure of a biological brain and aim to achieve cognitive performance comparable to that of humans but in a much shorter time.

These technologies now form the basis of machine learning and artificial intelligence systems that can perceive the environment and adapt their own behavior by analyzing the effects of previous actions and working autonomously. They are used in many areas of application, such as speech and image recognition and synthesis, autonomous driving and augmented reality systems, bioinformatics, genetic and molecular sequencing, and high-performance computing technologies.

Compared to conventional computing approaches, in order to perform complex functions, neural networks need to be initially "trained" with a large amount of known information that the network then uses to adapt by learning from experience. Training is an extremely energy-intensive process and as computing power increases, the neural networks' consumption grows very rapidly, doubling every six months or so.

Photonic circuits are a very promising technology for neural networks because they make it possible to build energy-efficient computing units. For years, the Politecnico di Milano has been working on developing programmable photonic processors integrated on silicon microchips only a few mm2 in size for use in the field of data transmission and processing, and now these devices are being used to build photonic neural networks.

"An artificial neuron, like a biological neuron, must perform very simple mathematical operations, such as addition and multiplication, but in a neural network consisting of many densely interconnected neurons, the energy cost of these operations grows exponentially and quickly becomes prohibitive. Our chip incorporates a photonic accelerator that allows calculations to be carried out very quickly and efficiently, using a programmable grid of silicon interferometers. The calculation time is equal to the transit time of light in a chip a few millimeters in size, so we are talking about less than a billionth of a second (0.1 nanoseconds)," says Francesco Morichetti, Head of the Photonic Devices Lab of the Politecnico di Milano.

"The advantages of photonic neural networks have long been known, but one of the missing pieces to fully exploit their potential was network training.. It is like having a powerful calculator, but not knowing how to use it. In this study, we succeeded in implementing training strategies for photonic neurons similar to those used for conventional neural networks. The photonic 'brain' learns quickly and accurately and can achieve precision comparable to that of a conventional neural network, but faster and with considerable energy savings. These are all building blocks for artificial intelligence and quantum applications," adds Andrea Melloni, Director of Polifab the Politecnico di Milano micro and nanotechnology center.   ... ' 

Study; Neural Networks and the Brain

During the early days of neural nets in the 1980s, we brought in a neuroscientist to talk to our group.  He emphasized the great difference between the neurons of the brain and the statistical neural nets we were experimenting with then.   How much closer are we now?

Study Urges Caution When Comparing Neural Networks to the Brain  By MIT News, November 3, 2022

The researchers studied over 11,000 neural networks trained to simulate the function of grid cells (part of the brain's navigation system) and found that very specific constraints not found in biological systems are needed for neural networks to generate grid-cell-like activity.

Researcher Rylan Schaeffer said, "What this suggests is that in order to obtain a result with grid cells, the researchers training the models needed to bake in those results with specific, biologically implausible implementation choices."

The analysis showed that close to 90% of the neural networks successfully learned path integration (a prediction of an animal's next location based on a given starting point and velocity), but grid-cell-like activity patterns were produced by just 10% of the networks.... 

Neural networks form the basis of many artificial intelligence systems for applications such as speech recognition, computer vision, and medical image analysis.... '

Full Article.  

Turning Physical Systems into Neural Nets

 Replacing electronic processors with s purely physical system? Hmm ... considering the uses.  Worth  a look.  

Physical Systems Perform ML Computations, By Cornell Chronicle, January 31, 2022

Cornell University researchers have trained physical systems to execute generic machine learning computations, demonstrating an early but viable substitute for conventional electronic processors.

The training process enabled demonstrations with mechanical, optical, and electrical physical systems.

The mechanical system involved a titanium plate positioned atop a speaker to create a driven multimode mechanical oscillator; the optical system beamed a laser through a nonlinear crystal to convert the incoming light's colors into new colors by combining photon pairs, and the electrical system harnessed an electronic circuit with a resistor, a capacitor, an inductor, and a transistor.

The researchers fed each system pixels of an image of a handwritten number, encoded in a light pulse or an electrical voltage, and returned a similar type of optical pulse or voltage as output.

"It turns out you can turn pretty much any physical system into a neural network," said Cornell's Peter McMahon.

From Cornell Chronicle

View Full Article    

Better Detection of Earthquakes with Machine Learning

New work in the space using Machine Learning, we had proposed related methods when studying neural models.  Here a convolutional neural network.

Researchers Create Earthquake System Model with Better Detection Capabilities  in CACM

University of Wyoming,  October 12, 2021

The University of Wyoming's Pejman Tahmasebi and Tao Bai have invented a machine learning (ML) model that boosts the accuracy of earthquake detection significantly over current models. Tahmasebi said the model processes signal data recorded by seismometers, and can automatically distinguish seismic events from seismic noise. The model combines existing long short-term memory and fully convolutional network ML models; the former captures data signal changes over time, and the latter filters out hidden features of seismic events. Tahmasebi said the model boasts 89.1% classification accuracy, a 14.5% improvement over the state-of-the-art ConvNetQuake model.

Full article

The Risk and Uncertainty of it All

A space we played in early on, determining measures of uncertainty as we built AI and expert system based models.   Its part of any decision driving system,  That includes measures of uncertainty and risk.  Glad to see efforts to include it in machine learning systems today.  Such measures should lead to model improvements.  Now can these models also learn what aspects of the model design and data directly create problems of certainty?   

A Neural Network Learns When It Should Not Be Trusted    By MIT News

 

The advance could enhance safety and efficiency in artificial intelligence-assisted decision-making. 

Massachusetts Institute of Technology researchers have developed a way for deep learning neural networks to rapidly estimate confidence levels in their output.

Researchers at the Massachusetts Institute of Technology (MIT) and Harvard University have enabled a neural network to rapidly process data, yielding both predictions and confidence levels based on the quality of the available data.

This deep evidential regression technique, which estimates uncertainty from a single run of the neural network, could lead to safer results.

The team designed the network with bulked-up output, generating not only a decision but also a new probabilistic distribution capturing the evidence supporting that decision; these evidential distributions directly capture the model's confidence in its forecast.

Included is any uncertainty within the underlying input data and the model's final decision, which indicates whether uncertainty can be reduced by modifying the network itself, or whether the input data is merely noisy.

MIT's Daniela Rus said, "By estimating the uncertainty of a learned model, we also learn how much error to expect from the model, and what missing data could improve the model."

From MIT News

Illusory Perceptions

 In our early work in this area we thought we found such 'illusions'. Based on the text here, these were not the same thing as mentioned here,  but we named them 'illusions', inspired by thoughts of biomimicry.  Is this a hint we are getting closer to brain models?

AI Also Has Illusory Perceptions

RUVID/Network of Valencian Universities for the Promotion of Research, Development, and Innovation

October 16, 2020

Researchers at Spain’s Universitat de València (UV) and Pompeu Fabra University have found that convolutional neural networks (CNN) are affected by visual illusions, much like the human brain. The researchers trained CNNs for simple tasks and found they were susceptible to visual illusions of brightness, although the illusions may not coincide with biological illusory perceptions. Said UV's Jesús Malo, "This is one of the factors that leads us to think that it is not possible to establish analogies between the simple concatenation of artificial neural networks and the much more complex human brain." The researchers warned in a separate study about the use of CNNs to study human vision. Said Malo, "In addition to the intrinsic limitations of these artificial networks to model vision, the non-linear behavior of flexible architectures can be very different from that of the biological visual system."

Net Model Finds Small Objects in Dense Images

This sounds like a very interesting problem in general. Say looking for tiny errors in manufacturing.   And a number of others.

NIST Neural Network Model Finds Small Objects in Dense Images
NIST
August 4, 2020

Computer scientists at the U.S. National Institute of Standards and Technology (NIST) have created a neural network model to detect small geometric objects in dense images. The researchers modified a network architecture developed by German scientists for analyzing biomedical images in order to retrieve raw data from journal articles that had been degraded or otherwise lost. The images present data points with various markers, mainly circles, triangles, and squares, both filled and open, of differing size and clarity. The model captured 97% of objects in a defined set of test images, finding their centers to within a few pixels of manually selected sites. NIST's Adele Peskin said the technique could find use in other applications, because "object detection is used in a wide range of image analyses, self-driving cars, machine inspections, and so on, for which small, dense objects are particularly hard to locate and separate." .... ' 

Security of The Form and Parameters of Neural Nets.

Out of Cornell University an intriguing article that deals with how neural nets react to adversarial attacks in their energy consumption.   As predicted by simulation.   Akin to how you might test a system by giving it questions that you know would take time for a human to do, but are easy for machines.  Doing this repeatedly could reveal indications to the form and parameters of the network involved. Which contains the 'knowledge' involved.  Threats continue to get very innovative.

Sponge Examples: Energy-Latency Attacks on Neural Networks

By Ilia Shumailov, Yiren Zhao, Daniel Bates, Nicolas Papernot, Robert Mullins, Ross Anderson

The high energy costs of neural network training and inference led to the use of acceleration hardware such as GPUs and TPUs. While this enabled us to train large-scale neural networks in datacenters and deploy them on edge devices, the focus so far is on average-case performance. In this work, we introduce a novel threat vector against neural networks whose energy consumption or decision latency are critical. We show how adversaries can exploit carefully crafted sponge examples, which are inputs designed to maximise energy consumption and latency.

We mount two variants of this attack on established vision and language models, increasing energy consumption by a factor of 10 to 200. Our attacks can also be used to delay decisions where a network has critical real-time performance, such as in perception for autonomous vehicles. We demonstrate the portability of our malicious inputs across CPUs and a variety of hardware accelerator chips including GPUs, and an ASIC simulator. We conclude by proposing a defense strategy which mitigates our attack by shifting the analysis of energy consumption in hardware from an average-case to a worst-case perspective.  ... "

Also being discussed at Schneier, where there is some interesting comment going on.