Good to see us listed as partners on the Neurobus website. Looks like we may already be embedded in some of their products..
Thanks GS,
... and their other industry partners are Intel and Prophesee.
Prophesee's DVS can instantaneously detect changes above a set threshold in a field of view. This is another project on which Brainchip and Prophesee are working together.
Hi Dio,
A little left field but can you see a way for the JAST learning rules to be implemented in to TENNs? (orthogonal polynomials)
From memory the JAST rules only took up something like 64K lines of code.
Way, way, way out side of my pay grade.
Well, until you asked the question, I hadn't thought about it.
In my mind, I had considered the division of labour as being Akida doing the inference/classification and TENNS handling the temporal element. But we have learnt that TENNS can be use without Akida, so there's another neat hypothesis down the plughole.
I haven't really got into how TENNS works, but we do know that, unlike Akida, it uses MACs. However it also uses the same sparse weights and activations.
I'm more a pictures person than a words person and I'm not personally acquainted with converging orthogonal polynomials, but some of my best friends are.
This is an abstract from the TENNS patent:
WO2023250093A1 METHOD AND SYSTEM FOR IMPLEMENTING TEMPORAL CONVOLUTION IN SPATIOTEMPORAL NEURAL NETWORKS 20220622
COENEN OLIVIER JEAN-MARIE DOMINIQUE [US]; PEI YAN RU [US]
[0012] A
ccording to an embodiment of the present disclosure, disclosed herein is a neural network system that includes an input interface, a memory including a plurality of temporal and spatial layers, and a processor.
The input interface is configured to receive sequential data that includes temporal data sequences. The memory is configured to store a plurality of group of first temporal kernel values, a first plurality of First-In-FirstOut (FIFO) buffers corresponding to a current temporal layer.
The memory further implements a neural network that includes a first plurality of neurons for the current temporal layer, a corresponding group among the plurality of groups of the first temporal kernel values is associated with each connection of a corresponding neuron of the first plurality of neurons.
The processor is configured to allocate the first plurality of FIFO buffers to a first group of neurons among the first plurality of neurons.
The processor is then configured to receive a first temporal sequence of the corresponding temporal data sequences into the first plurality of FIFO buffers allocated to the first group of neurons from corresponding temporal data sequences over a first time window.
Thereafter, the processor is configured to perform, for each connection of a corresponding neuron of the first group of neurons, a first dot product of the first temporal sequence of the corresponding temporal data sequences within a corresponding FIFO buffer of first plurality of FIFO buffers with a corresponding temporal kernel value among the corresponding group of the first temporal kernel values.
The corresponding temporal kernel values are associated with a corresponding connection of the corresponding neuron of the first group of neurons.
The processor is then further configured to determine a corresponding potential value for the corresponding neurons of the first group of neurons based on the performed first dot product and then generates a first output response based on the determined corresponding potential values.
Th
e TENNs as disclosed herein may effectively learn both spatial and temporal correlations from the input data.
[0050] According to an embodiment, the spatiotemporal networks may be configured to perform the temporal convolution operations either in a buffered temporal convolution mode or a recurrent temporal convolution mode, and may be alternatively referred to as a “buffer mode” or a “recurrent mode”, respectively.
[0051] According to an embodiment, the spatiotemporal network may be configured with a plurality of spatiotemporal convolution layers. Each of the spatiotemporal layers may be further split into plurality of temporal and spatial convolution layers. The kernels for the temporal and spatial convolution layers are represented as a sum over a set of basis functions, such as orthogonal polynomials, where the coefficients of the basis functions are trainable parameters of the network. This basis function representation compresses the number of parameters of the spatiotemporal network, which makes the training of the spatiotemporal network stable and resistant to overfitting.*
* Overfitting is responsible for the "hallucinations" experienced on OpenAI.
The JAST patent summarizes the rules as follows:
US11853862B2 Method, digital electronic circuit and system for unsupervised detection of repeating patterns in a series of events 20161121
- Input events (“spikes”) are grouped into fixed-size packets. The temporal order between events of a same packet is lost, which may seem a drawback, but indeed increases robustness as it eases the detection of distorted patterns and makes the method insensitive to changes of the event rate.
- Weighted or un-weighted synapses are replaced by a set of binary weights. Learning only requires flipping some of these binary weights and performing sums and comparisons, thus minimizing the computational burden.
- The number of binary weights which is set to “1” for each neuron does not vary during the learning process. This avoids ending up with non-selective or non-sensible neurons.
In other words, the synspses are GO-NO GO gates, depending on how they are programmed from the model.
[0062] T
he neural processor 320 may correspond to a neural processing unit (NPU). The (NPU) is a specialized circuit that implements all the necessary control and arithmetic logic necessary to execute machine learning algorithms, typically by operating on models such as artificial neural networks (ANNs) and spiking neural networks (SNNs).
However, does the loss of temporal order
within a packet exclude TENNS? I wouldn't think so, as the temporal order
between the packets is retained.
So the question is
A - whether the JAST rules are compatible with TENNS COPs?; and
B- whether there is any advantage in combining them?
The short answer is: the chooks ate my homework.
Truth is I haven't dug in depth into the TENNS patent or looked at the pictures.
So I think the old engineering adage "If it ain't broke, ..." applies.
As my mother used to say "Leave it alone - you'll make it explode!"
www.newelectronics.co.uk
