Deterministic Sequence Stochastic Computing for Neural Network Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Stochastic computing faces challenges due to random fluctuation errors and area overhead associated with linear feedback shift registers, which increase latency and power consumption as sequence lengths increase, limiting its practical application.
Innovation Solution
The use of deterministic sequences instead of pseudo-random sequences for stochastic computing, allowing for guaranteed error bounds and reduced sequence lengths, thereby simplifying hardware and software requirements for neural network inference computations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If pseudo-random sequences are used for stochastic computing, then fault tolerance and simple hardware implementation are achieved, but random fluctuation errors increase and area overhead increases due to linear feedback shift registers
Solution Approach 1:
The patent uses deterministic sequences that replicate the statistical properties of pseudo-random sequences without requiring LFSR hardware. The deterministic sequences are generated through simple repetition and reversal of base sequences, copying the desired randomness characteristics while eliminating the source of random fluctuation errors.
Solution Approach 2:
The patent replaces the mechanical LFSR system with a deterministic sequence generation approach. Instead of using linear feedback shift registers to generate pseudo-random sequences, the invention uses simple deterministic sequences generated through repetition and reversal operations, eliminating the need for complex feedback mechanisms.
2Measurement precision
If sequence lengths are increased to reduce random fluctuation errors, then measurement precision improves, but area overhead and latency increase due to linear feedback shift registers
Solution Approach 1:
The patent changes the fundamental parameter of sequence generation from pseudo-random to deterministic. This allows for shorter sequence lengths to achieve the same computation accuracy because deterministic sequences eliminate random fluctuation errors. The parameter change from LFSR-based generation to simple repetition/reversal generation dramatically reduces the required sequence length.
3Measurement precision
If sequence lengths are increased to reduce random fluctuation errors, then measurement precision improves, but area overhead increases due to linear feedback shift registers
Solution Approach 1:
The patent copies the statistical properties of pseudo-random sequences using deterministic sequences generated through simple repetition and reversal. This copying approach achieves the desired computation accuracy without requiring large-area LFSR hardware, as the deterministic sequences can be generated with minimal hardware overhead.
Solution Approach 2:
The patent extracts only the essential statistical properties needed for stochastic computing from pseudo-random sequences, discarding the need for LFSR hardware. By taking out only the necessary randomness characteristics and implementing them through deterministic sequences, the invention eliminates the area overhead associated with full pseudo-random sequence generation.
4Reliability
If pseudo-random sequences are used for stochastic computing, then fault tolerance is achieved, but device complexity increases due to linear feedback shift registers
Solution Approach 1:
The patent substitutes the complex LFSR mechanical system with a simple deterministic sequence generation approach. Instead of implementing feedback logic and shift register hardware, the invention uses straightforward repetition and reversal operations that can be implemented with minimal hardware, thereby reducing device complexity while maintaining fault tolerance.
Solution Approach 2:
The patent copies the fault tolerance property of pseudo-random sequences using deterministic sequences. By replicating the statistical characteristics that provide fault tolerance without copying the complex LFSR hardware, the invention achieves reliability with simpler device architecture.
Data Source
AI summary
Providing neural network output values by receiving a neural network having defined node weights, determining network activations according to network input data, for each activation x of the network activations and each node weight w of the defined node weights: generating an activation sequence which is a unary representation of x/nx, a first weight sequence WL, and a second weight sequence WR, where the sequence W=(Wn<sub2>w</sub2>, Wn<sub2>w</sub2>−1, . . . , W2, W1, W0) is a unary representation of w/nw, the sequence V=(w0, W1, W2, . . . , Wn<sub2>w</sub2>−1, Wn<sub2>w</sub2>) is the reverse of W, and nx is a multiple of nw so that WL and WR are of length nx; and computing MS(X, WL)+MS(X, WR) as an approximation for 2(x/nx)(w/nw), wherein MS(.,.) denotes stochastic multiplication, and providing a neural network output based on, at least in part, a set of computed approximations for 2(x/nx)(w/nw) values.