-
Unlocking the potential of two-point cells for energy-efficient and resilient training of deep nets
Authors:
Ahsan Adeel,
Adewale Adetomi,
Khubaib Ahmed,
Amir Hussain,
Tughrul Arslan,
W. A. Phillips
Abstract:
Context-sensitive two-point layer 5 pyramidal cells (L5PCs) were discovered as long ago as 1999. However, the potential of this discovery to provide useful neural computation has yet to be demonstrated. Here we show for the first time how a transformative L5PCs-driven deep neural network (DNN), termed the multisensory cooperative computing (MCC) architecture, can effectively process large amounts…
▽ More
Context-sensitive two-point layer 5 pyramidal cells (L5PCs) were discovered as long ago as 1999. However, the potential of this discovery to provide useful neural computation has yet to be demonstrated. Here we show for the first time how a transformative L5PCs-driven deep neural network (DNN), termed the multisensory cooperative computing (MCC) architecture, can effectively process large amounts of heterogeneous real-world audio-visual (AV) data, using far less energy compared to best available 'point' neuron-driven DNNs. A novel highly-distributed parallel implementation on a Xilinx UltraScale+ MPSoC device estimates energy savings up to 245759 $ \times $ 50000 $μ$J (i.e., 62% less than the baseline model in a semi-supervised learning setup) where a single synapse consumes $8e^{-5}μ$J. In a supervised learning setup, the energy-saving can potentially reach up to 1250x less (per feedforward transmission) than the baseline model. The significantly reduced neural activity in MCC leads to inherently fast learning and resilience against sudden neural damage. This remarkable performance in pilot experiments demonstrates the embodied neuromorphic intelligence of our proposed cooperative L5PC that receives input from diverse neighbouring neurons as context to amplify the transmission of most salient and relevant information for onward transmission, from overwhelmingly large multimodal information utilised at the early stages of on-chip training. Our proposed approach opens new cross-disciplinary avenues for future on-chip DNN training implementations and posits a radical shift in current neuromorphic computing paradigms.
△ Less
Submitted 22 December, 2022; v1 submitted 24 October, 2022;
originally announced November 2022.
-
A comparison of partial information decompositions using data from real and simulated layer 5b pyramidal cells
Authors:
Jim W. Kay,
Jan M. Schulz,
W. A. Phillips
Abstract:
Partial information decomposition allows the joint mutual information between an output and a set of inputs to be divided into components that are synergistic or shared or unique to each input. We consider five different decompositions and compare their results on data from layer 5b pyramidal cells in two different studies. The first study was of the amplification of somatic action potential outpu…
▽ More
Partial information decomposition allows the joint mutual information between an output and a set of inputs to be divided into components that are synergistic or shared or unique to each input. We consider five different decompositions and compare their results on data from layer 5b pyramidal cells in two different studies. The first study was of the amplification of somatic action potential output by apical dendritic input and its regulation by dendritic inhibition. We find that two of the decompositions produce much larger estimates of synergy and shared information than the others, as well as large levels of unique misinformation. When within-neuron differences in the components are examined, the five methods produce more similar results for all but the shared information component, for which two methods produce a different statistical conclusion from the others. There are some differences in the expression of unique information asymmetry among the methods. It is significantly larger, on average, under dendritic inhibition. Three of the methods support a previous conclusion that apical amplification is reduced by dendritic inhibition. The second study used a detailed compartmental model to produce action potentials for many combinations of the numbers of basal and apical synaptic inputs. Two analyses of decompositions are conducted on subsets of the data. In the first, the decompositions reveal a bifurcation in unique information asymmetry. For three of the methods this suggests that apical drive switches to basal drive as the strength of the basal input increases, while the other two show changing mixtures of information and misinformation. Decompositions produced using the second set of subsets show that all five decompositions provide support for properties of cooperative context-sensitivity - to varying extents.
△ Less
Submitted 13 June, 2022;
originally announced June 2022.
-
Contrasting information theoretic decompositions of modulatory and arithmetic interactions in neural information processing systems
Authors:
Jim W. Kay,
William A. Phillips
Abstract:
Biological and artificial neural systems are composed of many local processors, and their capabilities depend upon the transfer function that relates each local processor's outputs to its inputs. This paper uses a recent advance in the foundations of information theory to study the properties of local processors that use contextual input to amplify or attenuate transmission of information about th…
▽ More
Biological and artificial neural systems are composed of many local processors, and their capabilities depend upon the transfer function that relates each local processor's outputs to its inputs. This paper uses a recent advance in the foundations of information theory to study the properties of local processors that use contextual input to amplify or attenuate transmission of information about their driving inputs. This advance enables the information transmitted by processors with two distinct inputs to be decomposed into those components unique to each input, that shared between the two inputs, and that which depends on both though it is in neither, i.e. synergy. The decompositions that we report here show that contextual modulation has information processing properties that contrast with those of all four simple arithmetic operators, that it can take various forms, and that the form used in our previous studies of artificial neural nets composed of local processors with both driving and contextual inputs is particularly well-suited to provide the distinctive capabilities of contextual modulation under a wide range of conditions. We argue that the decompositions reported here could be compared with those obtained from empirical neurobiological and psychophysical data under conditions thought to reflect contextual modulation. That would then shed new light on the underlying processes involved. Finally, we suggest that such decompositions could aid the design of context-sensitive machine learning algorithms.
△ Less
Submitted 15 March, 2018;
originally announced March 2018.
-
Partial Information Decomposition as a Unified Approach to the Specification of Neural Goal Functions
Authors:
Michael Wibral,
Viola Priesemann,
Jim W. Kay,
Joseph T. Lizier,
William A. Phillips
Abstract:
In many neural systems anatomical motifs are present repeatedly, but despite their structural similarity they can serve very different tasks. A prime example for such a motif is the canonical microcircuit of six-layered neo-cortex, which is repeated across cortical areas, and is involved in a number of different tasks (e.g.sensory, cognitive, or motor tasks). This observation has spawned interest…
▽ More
In many neural systems anatomical motifs are present repeatedly, but despite their structural similarity they can serve very different tasks. A prime example for such a motif is the canonical microcircuit of six-layered neo-cortex, which is repeated across cortical areas, and is involved in a number of different tasks (e.g.sensory, cognitive, or motor tasks). This observation has spawned interest in finding a common underlying principle, a 'goal function', of information processing implemented in this structure. By definition such a goal function, if universal, cannot be cast in processing-domain specific language (e.g. 'edge filtering', 'working memory'). Thus, to formulate such a principle, we have to use a domain-independent framework. Information theory offers such a framework. However, while the classical framework of information theory focuses on the relation between one input and one output (Shannon's mutual information), we argue that neural information processing crucially depends on the combination of \textit{multiple} inputs to create the output of a processor. To account for this, we use a very recent extension of Shannon Information theory, called partial information decomposition (PID). PID allows to quantify the information that several inputs provide individually (unique information), redundantly (shared information) or only jointly (synergistic information) about the output. First, we review the framework of PID. Then we apply it to reevaluate and analyze several earlier proposals of information theoretic neural goal functions (predictive coding, infomax, coherent infomax, efficient coding). We find that PID allows to compare these goal functions in a common framework, and also provides a versatile approach to design new goal functions from first principles. Building on this, we design and analyze a novel goal function, called 'coding with synergy'. [...]
△ Less
Submitted 3 October, 2015;
originally announced October 2015.
