The Week in AI is a roundup of high-impact AI/ML research and news to keep you up to date in the fast-moving world of enterprise machine learning. From a reconfigurable, LEGO-like neural network chip design to an AI system making new human DNA discoveries, here are this week’s highlights.
MIT engineers have created a stackable, reconfigurable AI chip. Its alternating layers of sensing and processing elements can communicate with one another using LEDs. This is a step toward modular vision, where LEGO-like chipware designs with unlimited expandability could keep devices up to date while reducing electronic waste.
Other modular chip designs use conventional wiring, which is difficult to sever and rewire for reconfiguration, to relay signals among layers. The updated design, as published in Nature Electronics, uses light rather than physical wires to transmit information through the chip. During reconfiguration, such as adding new light, pressure, or smell sensors or when updating processors, layers can be swapped out or stacked on.
The design is configured to perform basic image-recognition tasks by layering sensors, LEDs, and processors made from artificial synapses—arrays of memory resistors, or memristors, which function as a brain on a chip. Each array can be trained to process and classify signals directly on a chip without the need for external software or an Internet connection.
The chip’s computing core measures 4 square millimeters, or about the size of a piece of confetti. Equipped with three image-recognition “blocks”—each with an image sensor, optical communication layer, and artificial synapse array for classifying one of three letters, M, I, or T—the chip can correctly classify clear images but struggles to distinguish between blurry ones.
To address this, the team replaced the processing layer with a better “denoising” processor, allowing the chip to accurately identify lower-quality images. In the near future, the researchers plan to add more sensing and processing capabilities to the chip, which can be used in numerous applications. One example is modular chips built into electronics that consumers can use to add, say, image or voice recognition.
A group of Cornell computer scientists successfully developed an ML technique to analyze massive amounts of data from the quantum pyrochlore oxide metal Cd2Re2O7. This not only settles a debate about the material, but it sets the stage for future ML-aided insight about new phases of matter. The paper, “Harnessing Interpretable and Unsupervised Machine Learning to Address Big Data from Modern X-ray Diffraction,” was published in the Proceedings of the National Academy of Sciences.
The unsupervised, interpretable ML algorithm, called “X-ray powder diffraction (XRD) temperature clustering,” or X-TEC, can analyze 8TB of X-ray data in minutes to investigate key elements of Cd2Re207. During their analysis, researchers discovered important insights into electron behavior in the material, detecting what is known as the pseudo-Goldstone mode.
Pseudo-Goldstone allows researchers to understand how atoms and electrons position themselves in an orderly fashion to optimize the interaction within the astronomically large “community” of electrons and atoms.
The creation of X-TEC is significant for three reasons. First, it shows that ML can be used to analyze voluminous XRD data. Second, it settles a debate concerning the physics of Cd2Re207. Third, it showcases what collaboration between physicists and computer scientists can accomplish.
The researchers hope that leveraging ML is the first step in being able to answer key challenging scientific questions accompanying any new discovery of new phases of quantum matter.
A team of scientists at the Max Delbruck Center for Molecular Medicine in the Helmholtz Association (MDC) developed an ML program called “ikarus” to reliably distinguish cancer cells from healthy ones.
Ikarus found a pattern in tumor cells that’s common in different types of cancer, consisting of a characteristic combination of genes. According to the team’s paper in the journal Genome Biology, the algorithm also detected types of genes in the pattern that had never been clearly linked to cancer before.
The team used data from lung and colorectal cancer cells to train the algorithm before applying it to datasets of other kinds of tumors. During the training phase, ikarus had to find a list of characteristic genes that it then used to categorize the cells as cancerous or noncancerous. In the postlearning phase, the algorithm reliably categorized cells from other types of cancer, such as in tissue samples from liver cancer or neuroblastoma patients, at a very high success rate.
In hospitals, pathologists tend to examine tissue samples of tumors only under the microscope to identify the various cell types, which is a laborious, time-consuming process. While the team believes this step could one day become a fully automated process using ikarus, they also hope that the project will go far beyond the classification of healthy versus cancerous cells, including helping doctors identify the best cancer therapy for each patient.
Scientists from DeepMind and the University of Washington (UW) released a new AI algorithm that deciphered the structure at the heart of inheritance. This structure was a massive complex of roughly 1,000 proteins called nuclear pore complexes (NPCs), which helps channel DNA to the rest of the cell. The AI model is built on top of AlphaFold and RoseTTAfold, two predecessor models released to the public by DeepMind and David Baker’s lab at UW, respectively.
With a donut-shaped architecture, NPCs are among the largest protein complexes in the human body and strictly monitor the ins and outs of molecular messengers. NPCs are essential for gene therapy, mRNA-type vaccines, and potentially other genetic treatments researchers haven’t yet imagined.
Traditionally, scientists have sought to decode NPC structure using biochemical methods to tamper with its normal function or X-rays to scan its crystalline structure. Both methods are painstakingly time-consuming.
To improve this process with AI, researchers tackled two issues that have hindered the progress of cellular structure discovery: the lack of high-quality datasets and the lack of appropriate computing methods to analyze these massively large datasets.
Despite promising results when compared to previous methods, the scientists warn that AI isn’t a panacea. Challenges remain regarding the ability to analyze NPCs due to their capacity to change configuration according to their environment—for example, to widen their channels when nested inside the nuclear envelope.
LEGO-like neural network chip designs are proof that the world is moving toward a more sustainable future, where cellphones, smartwatches, and other wearable devices don’t have to be shelved or discarded for a newer model. We are already seeing the modular framework happen in the car manufacturing industry.
As for the advancement of biomedical technology, researchers count heavily on AI systems such as ikarus to help identify genes that are potential drivers of cancer. A better understanding of a tumor’s immediate environment, thanks to AI, can help doctors select the best therapy for their patients.
Until next time, stay informed and get involved!