Related to the video posted earlier but now posted by Edge Impulse on their LinkedIn page:
BRN Discussion Ongoing
- Thread starter TechGirl
- Start date
Related to the video posted earlier but now posted by Edge Impulse on their LinkedIn page:
Fullmoonfever
Top 20
Just googling for dots and this presso / paper I posted about back in Jan popped up.
Couldn't find it back then but it's HERE if anyone wanted a read.
@Diogenese thoughts whenever if you have time or anything of interest in it?
TIA
Hi Fmf,
The paper discusses a model (AkidaNet) used for distinguishing weed from crop and describes how it was adapted to run on Akida 1000 using CNN2SNN.
The model has sveral layers, the first using 8-bits to deal with the model, the remaining layers running on 4 bit weights and activations. This is interesting in that the recently publishes BRN patent application added ALUs to the NPU, presumably to handle 8-bits more efficiently, so these results would not reflect the performance of Akida 2, or whatever version has the ALUs.
. AkidaNet
In this paper, we propose a new lightweight and energy-efficient model and convert it to a Spiking Neural Network(SNN) for implementation on a NSoC hardware platform for weed identification. AkidaNet is built with reference to thewell-known MobileNetV1 and VGG-11 CNN architectures. The full architecture of AkidaNet has been developed to enhance power efficiency and reduce computational latency as shown in Fig.3. Specifically, the model begins with four3×3regular convolutional layers (Conv2D), followed by eleven depthwise separable convolutional layers (Separable-Conv2D), and ends with a softmax activation with 4 outputs for crops/weeds classification. As can be observed in Fig.4,the differences among the standard convolution block, theMobileNetV1 DSC block, and the AkidaNet DSC block are illustrated in detail. The AkidaNet structure allows the NSoC platform to process the AkidaNet DSC blocks more efficiently with less memory. Furthermore, it is important to note that instead of using max pooling operations in this model, a stride of 2 is used for the specific layers including Conv2D 0, Conv2D 2, SeparableConv2D 4, Separable-Conv2D 6 and SeparableConv2D 12 to reduce the resolution of the outputs of all standard convolution and AkidaNet DSC blocks. In addition, all standard convolution and depthwise separable convolution operations have a kernel size of 3×3and zero padding. Finally, the Classification top layer is kept as a separable convolutional layer instead of a dense layer to reduce memory usage and allow AkidaNet’s architecture to utilize the NSoC hardware more efficiently. With regard to the AKD1000 NSoC hardware version, it currently supports most layers found in feed-forward network architectures such as Dense, Standard Convolutional, Depth-wise Separable Convolutional, Batch Normalization, MaxPooling Layers. AkikaNet’s optimized architecture came from a limited architecture search that measured accuracy, power, and latency of each model variation deployed on the NSoC. Additionally, we also investigated the impacts of max pooling and stride-2 convolution on reducing spatial dimensions. As a result, the stride-2 option gave better performance on power and accuracy. The AkidaNet model can be adjusted to further reduce computational cost by scaling the number of output channels in each layer with a hyper-parameter α, or the width multiplier, inspired by the original MobileNet paper [43]. Specifically, the range of αis from 0 to 1, with typical settings of 0.25, 0.5,0.75 and 1. For example, if α= 1, it will be the baseline AkidaNet. If α= 0.5, it will generate a model with only half the output channels used in each layer. Decreasing the output channels results in the decreasing of the number of model parameters, which reduces the final size of the model. Therefore, changing the width multiplier αin AkidaNet allows one to meet the resource constraints of the NSoC platform for real-time weed recognition in practice with a trade off , accuracy, and model size.
D. AkidaNet to Spiking Neural Network Conversion
The goal of the AkidaNet design is allow target applications to achieve low latency and power consumption when implemented on the AKD1000 NSoC hardware platform. To optimize the performance of this model on edge devices, a potential solution mentioned in this paper is the use of Spiking Neural Networks (SNNs). According to [50, 51, 52], before converting the architecture of AkidaNet to SNN for execution on the AKD1000 NSoC, the model needs to be quantized to use 4-bit activations and 4-bit parameters. The quantization process plays an indispensable role in significantly reducing model size and power consumption when compared with CPU or GPU implementations. After that, the quantized AkidaNet model is converted to an SNN in a process called CNN2SNNas shown in Fig.5. Particularly, all the communication between neurons in the SNN takes the form of “spikes” or “events “corresponding to binary impulses that are generated when a neuron crosses a threshold level of activation. If all neurons on an Neural Processing Unit (NPU) do not cross the threshold ,it will generate no output. Hence, this feature is the key to the efficiency of SNNs and contributes to further minimizing computational cost.
E. Evaluation of AkidaNet and MobileNetV1 models onImageNet-1K dataset
We trained both MobileNetV1 and AkidaNet models from scratch on the ImageNet-1K classification dataset [54] using Tensorflow/Keras for 90 epochs on a single NVIDIA GeForceRTX 2080 Ti GPU with a batch size of 128 images. Weused Stochastic Gradient Descent (SGD) with momentumoptimizer and the learning rate was set to 0.1 for the first10 epochs and then decreased to 0.0001 with an exponentialdecay. We also used the L2 weight regularizer in Conv2Dkernels and in pointwise kernels of the SeparableConv2Dlayers. In addition, we applied basic data augmentation (i.e.,random resized cropping and horizontal flipping). We trainedat the resolution of 160x160 and evaluated at the resolutionof 224x224 based on [55]. Finally, the last step is Akidaconversion and evaluation. The model was quantized to 4bit weights and activations except for the first convolutionallayer’s weights, which were quantized to 8 bits. After that, themodel was tuned for an additional 10 epochs with an initiallearning rate of 0.0001 that was kept constant for 2 epochsand then decreased to 1e-8.
F. Evaluation of the AkidaNet model on the bccr-segsetcrop/weed image dataset
We performed extensive experiments to evaluate the performance of our proposed method and some baseline approaches. All experiments were executed on an Ubuntu machine with an Intel Core 3.7GHz i7 CPU, 32Gb RAM and a GeforceGTX 1080Ti GPU. Our network was implemented in Python using TensorFlow and the MetaTF ML framework [50]. We compared our model against the following models: ResNet-50,VGG-16, and InceptionV3. For all experiments, we used 80% (24,000 images) of the dataset for training the model and 20% (6,000 images) for testing. In order to have a fair comparison, all models were trained with an Adam optimizer, 10 epochs, a batch size 32, and a learning rate 0.0001. The input size of VGG-16, ResNet-50, AkidaNet α= 0.25, and AkidaNetα= 0.5was 224 ×224, while the input size of InceptionV3was 299 ×299. Moreover, All CNN models, including the AkidaNet variants, were pre-trained with the ImageNet-1kdataset and then trained on our bccr-segset dataset. In the next steps, we quantized the Keras AkidaNet model and then converted the resulting quantized model to an SNN for execution on the AKD1000 NSoC. Specifically, AkidaNet models with α= 0.25 and α= 0.5were quantized by using the cnn2snn.quantize function in MetaTF framework [50]. The first layer of weights were quantized to 8-bits while the remaining model weights and activations were quantized with 4-bits in order to meet the requirements of the NSoC hardware. After having obtained a quantized model with satisfactory performance, the Keras quantized model was converted into the SNN model by using the cnn2snn.convert function [50],which returns a model in an Akida-compatible format. This model file format can be run efficiently on the NSoC hardware device in inference mode.
cosors
👀
Totally off topic.
Who opens the 3000th page with her/his post?
"Hello Fellow Chippers,
Thanks to ZeeBot for creating this amazing new forum so we can all have civilized thoughtful BrainChip discussions.
I'll get us started, 2022 has already exceeded my expectations & we are only in the first few days of February.
Can't wait to see what the year brings.
Love all you guys & Girls xxxx"
Who opens the 3000th page with her/his post?
"Hello Fellow Chippers,
Thanks to ZeeBot for creating this amazing new forum so we can all have civilized thoughtful BrainChip discussions.
I'll get us started, 2022 has already exceeded my expectations & we are only in the first few days of February.
Can't wait to see what the year brings.
Love all you guys & Girls xxxx"
Last edited:
cosors
👀
Is this finally heading in the direction of MosChip?
Responsibilities
- Test plan, Test bench development, execution, and debugging
- Block and Chip level- IP/ASIC/SOC/CPU/AMS Verification
- SystemVerilog with Testbench methodologies UVM
- Verilog, VHDL, C++, Vera, e, System C
- Protocols: PCIe / USB / DDR / Ethernet / ARM
/CNN/RNN
- Scripting: Perl, Tcl, Unix sc
ripting
https://www.linkedin.com/jobs/view/...at-brainchip-3650966374/?originalSubdomain=in
ASIC Verification Engineer
BrainChip Hyderabad, Telangana, IndiaResponsibilities
- Test plan, Test bench development, execution, and debugging
- Block and Chip level- IP/ASIC/SOC/CPU/AMS Verification
- SystemVerilog with Testbench methodologies UVM
- Verilog, VHDL, C++, Vera, e, System C
- Protocols: PCIe / USB / DDR / Ethernet / ARM
/CNN/RNN- Scripting: Perl, Tcl, Unix sc
ripting
https://www.linkedin.com/jobs/view/...at-brainchip-3650966374/?originalSubdomain=in
Last edited:
Morning Chippers,
Ford article on self driving...
Having heard diddley from our company or Ford , one can only assume we are still engaged here ....?
Regards,
Esq.
Ford article on self driving...
Having heard diddley from our company or Ford , one can only assume we are still engaged here ....?
Regards,
Esq.
Tothemoon24
Top 20
Future Of Artificial Intelligence Dominated Air Combat Showcased In New Air Force Video
The Air Force details its growing efforts to bring artificial intelligence-driven air combat drones from a laboratory experiment to reality.
www.thedrive.com
The Air Force Research Laboratory (AFRL) posted the new video covering aspects of the Autonomous Aircraft Experimentation (AAx) initiative on the Defense Visual Information Distribution Service(DVIDS) website earlier today. AAx's main focus is on testing and refining artificial intelligence and machine learning-driven autonomous capabilities for use on future advanced uncrewed aircraft and helping to move those technologies out of the laboratory and onto actual operational platforms.
We are trying to figure out how to integrate artificially trained neural networks, trained in a simulation... into the real world," Bill "Evil" Gray, the chief test pilot at the Air Force's Test Pilot School, explains in the newly released video. "In this case [through AAx], integrate them into controlling an airplane."
"We need to recognize that AI [artificial intelligence] is here. It's here to stay. It's a powerful tool," Air Force Col. Tucker "Cinco" Hamilton, the service's chief of AI Test and Operations, says at another point in the footage. "Collaborative Combat Aircraft and that type of autonomy is revolutionary. And will be the future battle space."
Welcome to the Edge
www.zdnet.com
Ahead of AI, this other technology wave is sweeping in fast
Welcome to the edge, which is likely to shape and shift our jobs and businesses before AI makes its mark.
www.zdnet.com
Worker122
Regular
Close!3000? Just taking a pot shot at it....
Qualcomm and Sony Agree to Leverage Snapdragon to Power Smartphones
The continued partnership between Qualcomm and Sony, using Qualcomm’s Snapdragon Mobile Platform for Sony’s premium and mid-tier devices.
futurumgroup.com
Last edited:
Good morning,
Brainchip is once again getting "our" story out there for all interested parties to engage now or into the future, it's a case of
"the early bird catches the worm"....don't hesitate, spread the word believers, check out the latest marketing video presented
by Nandan in the link below.
Cheers Tech
Brainchip is once again getting "our" story out there for all interested parties to engage now or into the future, it's a case of
"the early bird catches the worm"....don't hesitate, spread the word believers, check out the latest marketing video presented
by Nandan in the link below.
Cheers Tech
Damo4
Regular
But is seems you were closer
Not if we all report a post on page 2999 and have it bump back
*no posts were harmed (reported) in the making of this milestone*
Edit: This 3000 is mine!
Last edited:
Yoda
Regular
Yes we can only hope that Akida 2.0 happens to be a great hook to catch some worms!Thanks for the reminder. " The early bird catches the worm"
Even though we've been riding this investment for what feels like some sort of time warp to the future and back to the past and then slingshotted back to the future again. Crazy bus driver is still waiting for more passengers.The bus is fully packed with us worms waiting to turn over a new leaf
View attachment 39386
Similar threads
