Optical neuromorphic processing with soliton crystal Kerr microcombs: Scaling the network in size and speed

Convolutional neural networks (CNNs), inspired by biological visual cortex systems, are a powerful category of artificial neural networks that can extract the hierarchical features of raw data to greatly reduce the network parametric complexity and enhance the predicting accuracy. They are of significant interest for machine learning tasks such as computer vision, speech recognition, playing board games and medical diagnosis [1-7]. Optical neural networks offer the promise of dramatically accelerating computing speed to overcome the inherent bandwidth bottleneck of electronics. Here, we demonstrate a universal optical vector convolutional accelerator operating beyond 10 Tera-OPS (TOPS - operations per second), generating convolutions of images of 250,000 pixels with 8-bit resolution for 10 kernels simultaneously — enough for facial image recognition. We then use the same hardware to sequentially form a deep optical CNN with ten output neurons, achieving successful recognition of full 10 digits with 900 pixel handwritten digit images with 88% accuracy. Our results are based on simultaneously interleaving temporal, wavelength and spatial dimensions enabled by an integrated microcomb source. We show that this approach is scalable and trainable to much more complex networks for demanding applications such as unmanned vehicle and real-time video recognition.Keywords: Optical neural networks, neuromorphic processor, microcomb, convolutional accelerator

Unsupervised Character Recognition with Graphene Memristive Synapses

Memristive devices being applied in neuromorphic computing are envisioned to significantly improve the power consumption and speed of future computing platforms. The materials used to fabricate such devices will play a significant role in their viability. Graphene is a promising material, with superb electrical properties and the ability to be produced sustainably. In this paper, we demonstrate that a fabricated graphene-pentacene memristive device can be used as synapses within Spiking Neural Networks (SNNs) to realise Spike Timing Dependent Plasticity (STDP) for unsupervised learning in an efficient manner. Specifically, we verify operation of two SNN architectures tasked for single digit (0-9) classification: (i) a simple single-layer network, where inputs are presented in 5x5 pixel resolution, and (ii) a larger network capable of classifying the Modified National Institute of Standards and Technology (MNIST) dataset, where inputs are presented in 28x28 pixel resolution. Final results demonstrate that for 100 output neurons, after one training epoch, a test set accuracy of up to 86% can be achieved, which is higher than prior art using the same number of output neurons. We attribute this performance improvement to homeostatic plasticity dynamics that we used to alter the threshold of neurons during training. Our work presents the first investigation of the use of green-fabricated graphene memristive devices to perform a complex pattern classification task. This can pave the way for future research in using graphene devices with memristive capabilities in neuromorphic computing architectures. In favour of reproducible research, we make our code and data publicly available https://anonymous.4open.science/r/c69ab2e2-b672-4ebd-b266-987ee1fd65e7.

Serotonin Signals Modulate Mushroom Body Output Neurons for Sustaining Water-Reward Long-Term Memory in Drosophila

Memory consolidation is a time-dependent process through which an unstable learned experience is transformed into a stable long-term memory; however, the circuit and molecular mechanisms underlying this process are poorly understood. The Drosophila mushroom body (MB) is a huge brain neuropil that plays a crucial role in olfactory memory. The MB neurons can be generally classified into three subsets: γ, αβ, and α′β′. Here, we report that water-reward long-term memory (wLTM) consolidation requires activity from α′β′-related mushroom body output neurons (MBONs) in a specific time window. wLTM consolidation requires neurotransmission in MBON-γ3β′1 during the 0–2 h period after training, and neurotransmission in MBON-α′2 is required during the 2–4 h period after training. Moreover, neurotransmission in MBON-α′1α′3 is required during the 0–4 h period after training. Intriguingly, blocking neurotransmission during consolidation or inhibiting serotonin biosynthesis in serotoninergic dorsal paired medial (DPM) neurons also disrupted the wLTM, suggesting that wLTM consolidation requires serotonin signals from DPM neurons. The GFP Reconstitution Across Synaptic Partners (GRASP) data showed the connectivity between DPM neurons and MBON-γ3β′1, MBON-α′2, and MBON-α′1α′3, and RNAi-mediated silencing of serotonin receptors in MBON-γ3β′1, MBON-α′2, or MBON-α′1α′3 disrupted wLTM. Taken together, our results suggest that serotonin released from DPM neurons modulates neuronal activity in MBON-γ3β′1, MBON-α′2, and MBON-α′1α′3 at specific time windows, which is critical for the consolidation of wLTM in Drosophila.

Circuits for integrating learned and innate valences in the insect brain

Animal behavior is shaped both by evolution and by individual experience. Parallel brain pathways encode innate and learned valences of cues, but the way in which they are integrated during action-selection is not well understood. We used electron microscopy to comprehensively map with synaptic resolution all neurons downstream of all Mushroom Body output neurons (encoding learned valences) and characterized their patterns of interaction with Lateral Horn neurons (encoding innate valences) in Drosophila larva. The connectome revealed multiple convergence neuron types that receive convergent Mushroom Body and Lateral Horn inputs. A subset of these receives excitatory input from positive-valence MB and LH pathways and inhibitory input from negative-valence MB pathways. We confirmed functional connectivity from LH and MB pathways and behavioral roles of two of these neurons. These neurons encode integrated odor value and bidirectionally regulate turning. Based on this we speculate that learning could potentially skew the balance of excitation and inhibition onto these neurons and thereby modulate turning. Together, our study provides insights into the circuits that integrate learned and innate to modify behavior.

Mesodiencephalic junction Gabaergic inputs are processed separately from motor cortical inputs in the basilar pons

The basilar pontine nuclei (bPN) receive inputs from the entire neocortex and constitute the main source of mossy fibers to the cerebellum. Despite their critical position in the cortico-cerebellar pathway, it remains unclear if and how the bPN process inputs. An important unresolved question is whether the bPN strictly receives excitatory inputs or also receives inhibitory inputs. In the present study, we identified the mesodiencephalic junction as a prominent source of GABAergic afferents to the bPN. We combined optogenetics and whole-cell patch clamp recordings and confirmed that the bPN indeed receives monosynaptic GABA inputs from this region. Furthermore, we found no evidence that these inhibitory inputs converge with motor cortex (M1) inputs at the single neuron level. We also found no evidence of any connectivity between bPN neurons, suggesting the absence of a local circuit. Finally, rabies tracings revealed that GABAergic MDJ neurons themselves receive prominent inputs from neocortical output neurons. Our data indicates that inhibition from the MDJ, and excitation from the neocortex remain separate streams of information through the bPN. It is therefore unlikely that inhibition in the bPN has a gating function, but rather shapes an appropriate output of the bPN during behavior.

Intrinsic maturation of sleep output neurons regulates sleep ontogeny in Drosophila

The maturation of sleep behavior across a lifespan (sleep ontogeny) is an evolutionarily conserved phenomenon. Mammalian studies have shown that in addition to increased sleep duration, early life sleep exhibits stark differences compared to mature sleep with regard to the amount of time spent in certain sleep states. How intrinsic maturation of sleep output circuits contributes to sleep ontogeny is poorly understood. The fruit fly Drosophila melanogaster exhibits multifaceted changes to sleep from juvenile to mature adulthood. Here, we use a non-invasive probabilistic approach to investigate changes in sleep architecture in juvenile and mature flies. Increased sleep in juvenile flies is driven primarily by a decreased probability of transitioning to wake, and characterized by more time in deeper sleep states. Functional manipulations of sleep-promoting neurons in the dFB suggest these neurons differentially regulate sleep in juvenile and mature flies. Transcriptomic analysis of dFB neurons at different ages and a subsequent RNAi screen implicate genes involved in distinct molecular processes in sleep control of juvenile and mature flies. These results reveal that dynamic transcriptional states of sleep output neurons contribute to changes in sleep across the lifespan.

A pair of dopamine neurons mediate chronic stress signals to induce learning deficit in Drosophila melanogaster

Chronic stress could induce severe cognitive impairments. Despite extensive investigations in mammalian models, the underlying mechanisms remain obscure. Here, we show that chronic stress could induce dramatic learning and memory deficits in Drosophila melanogaster. The chronic stress–induced learning deficit (CSLD) is long lasting and associated with other depression-like behaviors. We demonstrated that excessive dopaminergic activity provokes susceptibility to CSLD. Remarkably, a pair of PPL1-γ1pedc dopaminergic neurons that project to the mushroom body (MB) γ1pedc compartment play a key role in regulating susceptibility to CSLD so that stress-induced PPL1-γ1pedc hyperactivity facilitates the development of CSLD. Consistently, the mushroom body output neurons (MBON) of the γ1pedc compartment, MBON-γ1pedc>α/β neurons, are important for modulating susceptibility to CSLD. Imaging studies showed that dopaminergic activity is necessary to provoke the development of chronic stress–induced maladaptations in the MB network. Together, our data support that PPL1-γ1pedc mediates chronic stress signals to drive allostatic maladaptations in the MB network that lead to CSLD.

Synchronous spiking of cerebellar Purkinje cells during control of movements

The information that the brain transmits from one region to another is often viewed through the lens of firing rates. However, if the output neurons could vary the timing of their spikes with respect to each other, then through synchronization they could highlight information that may be critical for control of behavior. In the cerebellum, the computations that are performed by the cerebellar cortex are conveyed to the nuclei via inhibition. Yet, synchronous activity entrains nucleus neurons, making them fire. Does the cerebellar cortex rely on spike synchrony within populations of Purkinje cells (P-cells) to convey information to the nucleus? We recorded from multiple P-cells while marmosets performed saccadic eye movements and organized them into populations that shared a complex spike response to error. Before movement onset, P-cells transmitted information via a rate code: the simple spike firing rates predicted the direction and velocity of the impending saccade. However, during the saccade, the spikes became temporally aligned within the population, signaling when to stop the movement. Thus, the cerebellar cortex relies on spike synchronization within a population of P-cells, not individual firing rates, to convey to the nucleus when to stop a movement.

Bantam regulates the adult sleep circuit in Drosophila

Sleep is a highly conserved feature of animal life characterized by dramatic changes in behavior, neural physiology and gene expression. The gene regulatory factors responsible for these sleep-dependent changes remain largely unknown. microRNAs are post-transcriptional modulators of gene expression which have been implicated in sleep regulation. Our previous screen identified 25 sleep-regulating microRNAs in Drosophila melanogaster, including the developmental regulator bantam (ban). Here we show that ban promotes early nighttime sleep through a population of glutamatergic neurons- the γ5β′2a/β′2mp/β′2mp_bilateral Mushroom Body Output Neurons (MBONs). We found that knockdown of ban in these neurons led to a reduction in early night sleep. The γ5β′2a/β′2mp/β′2mp_bilateral MBONs were previously shown to be wake-promoting, suggesting that ban acts to inhibit these neurons. GCaMP calcium imaging revealed that bantam inhibits the neural activity of the γ5β′2a/β′2mp/β′2mp_bilateral MBONs during the night but not the day. Blocking synaptic transmission in the γ5β′2a/β′2mp/β′2mp_bilateral MBONs rescued the effect of ban knockdown on sleep. Together these results suggest that ban promotes night sleep via the inhibition of the γ5β′2a/β′2mp/β′2mp_bilateral MBONs. RNAseq further revealed that bantam negatively regulates the wake-promoting mRNAs Kelch and CCHamide-2 receptor in the γ5β′2a/β′2mp/β′2mp_bilateral MBONs. These experiments establish bantam as an active regulator of sleep and neural activity within the fly brain.

Photonic Reservoir Computer with Output Expansion for Unsupervized Parameter Drift Compensation

We present a method to improve the performance of a reservoir computer by keeping the reservoir fixed and increasing the number of output neurons. The additional neurons are nonlinear functions, typically chosen randomly, of the reservoir neurons. We demonstrate the interest of this expanded output layer on an experimental opto-electronic system subject to slow parameter drift which results in loss of performance. We can partially recover the lost performance by using the output layer expansion. The proposed scheme allows for a trade-off between performance gains and system complexity.