Deep Spreadsheets with ExcelNet

1 deep Spreadsheets with ExcelNetDavid FouheyCarnegie Mellon UniversityPittsburgh, MaturanaCarnegie Mellon UniversityPittsburgh, coding stuff/Commie open source license/Backpropagation,Synergy,Enterpri se ready,Turnkey solution,WYSIWYG Weight Editing,Table 1. ExcelNet versus lesser introduceEXCELNET, the premier solution for DeepLearning in try it today!Author KeywordsDeep; neural ; Excel; spreadsheet; Convolutional NeuralNetworkINTRODUCTIONAs anyone in machine learning and computer vision will tellyou, deep Learning is the right tool to solve the problem. Andas anyone in business and finance will tell you, Excel is theright platform to implement your solution. But until now, therehas been no way to do deep Learning in Excel. To fill thisgap we have developedEXCELNET, the ultimate synergy ofspreadsheets and deep neural the human brain, Excel has cells . In Excel, as in thebrain, these cells are organized in columns.

2 Our approachis to typeweightsinto those cells . We group theweightsinto layers , and put each layer in a sheet , which can beeasily previewed and modified by the user. We then leveragethe sophisticated multiplication and addition capabilities ofExcel to multiply and add theseweightsto perform advancedtasks such as recognizing handwritten digits [3].1Or LibreOffice, if you re an open source communist. Excel is atrademark of Microsoft, we summarized in Table 1, this approach has many advantagesover other complicated solutions like Caffe [2], Theano [1], orMatConvNet [4]. ExcelNet does not have backpropagation,but it is not needed, because the interface puts the weights atthe tip of the user s fingers! To help you get started though,we provide some initial weights using a strategy similar tothe pre-training process in common use: we build a CNNin MatConvNet [4] that recognizes digits and transfer theseweights to an Excel spreadsheet.

3 We then allocate activationmaps in other Spreadsheets and define these activations usingExcel s powerful functions. Once done, Excel performs allmemory management and advantage is that the power of deep Learning is now inyourhands. There is no need for worrying about CUDA orBLAS: Excel s proven and battle-tested numerical codebasehandles it for you! There is no need to worry about platforms:Excel provides the necessary abstractions, and you can seam-lessly transition between Windows, Linux, and Mac! Thisenables the ultimate in computer vision deployment flexibility:you can develop on a Mac, and deploy simultaneously to clientand cloud show screenshots ofEXCELNETin action in Figure 1(a) shows the i/o layer sheet. Here, you can input newdata for the CNN to care of the restand gives you the answer, both in terms of a final probabilityas well as the most likely answer.

4 Figure 1(b) shows extractingactivations. Typically, this requires fussing with file formats, , or writing to execute your vision for deep Learning without learningall this useless nerd re in POINTS OFEXCELNETEXCELNET offers the ultimate in deep Learning abilities. Wesee a number of unique numerical abstraction layer:Worried about the cor-rectness of your new app? Excel is the most trusted numer-ical computing platform: BLAS and ATLAS are used forfluid simulations and other underwater-basket-weaving-likeendeavours ; Excel is used for finance. By building on a solidfoundation, we provide unsurpassed guarantees in terms ofalgorithmic security:Worried about your autonomousdriving platform inEXCELNET getting into the wrong handswhen you leave your phone in a bar in SF? Don t sweat it EXCELNETis compatible with the most secure forms ofdocument protection from Microsoft.

5 (a)(b)Figure 1. EXCELNETin action: (a) the interface, whereyouare in control of what goes into the network and where the results can be obtained instantly;(b) examining an activation map. Typically getting access to the power of deep Learning requires hiring a nerd or learning obscure nerdy you in control, and lets anyone harness the power of deep Learning!Low cost of entry:Not willing to invest in pricey GPUs? with a basic laptop, and a copy of LibreOffice, you too canhop on the deep Learning bandwagon with programming required:Afraid of missing out of thebuilding the next big thing in deep Learning because youdo not know how to program? ExcelNet makes it easy byleveraging a basic level knowledge of MS data science:Instead of slurping heaps of datafrom the Internet, ExcelNet enables you to hand-enter bothweights and inputs, just like in the good old days. The resultingnumbers are more DOESEXCELNETWORK?

6 Machine learning practitioners may wonder: how doesEX-CELNET work?It s simple: it s a self-contained .xls file that contains a sheetfor I/O, and then sheets for the filters and biases of severallayers of a convolutional neural net. The remaining sheetsapply mean-subtraction, convolutions, rectified linear units(ReLUs), fully connected layers, and a softmax function. Allof these operations can implemented using Excel I/O sheet then references the final output of the DO I USEEXCELNETTO CLASSIFY NEW DATA?Simply enter in your data in the rows and columns of the data-entry matrix, and watch the posterior distribution change inreal-time2! You may find it helpful to use the multiplier cellat the bottom that multiplies your entered numbers by a fixedvalue so you can type in 1 s and 0 s as opposed to 0 s-255 you have your prediction save , whichwill save all the activations and the current least if you re having a branded experience and using real MsEx-cel: LibreOffice is definitely try ExcelNet today!

7 REFERENCES1. James Bergstra, Olivier Breuleux, Fr d ric Bastien,Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins,Joseph Turian, David Warde-Farley, and Yoshua Theano: a CPU and GPU Math ExpressionCompiler. InProceedings of the Python for ScientificComputing Conference (SciPy).2. Yangqing Jia, Evan Shelhamer, Jeff Donahue, SergeyKarayev, Jonathan Long, Ross Girshick, SergioGuadarrama, and Trevor Darrell. 2014. Caffe:Convolutional Architecture for Fast Feature preprint (2014).3. Yann LeCun, L on Bottou, Yoshua Bengio, and PatrickHaffner. 2001. Gradient-Based Learning Applied toDocument Recognition. InIntelligent Signal Processing,S. Haykin and B. Kosko (Eds.). IEEE Press, 306 A. Vedaldi and K. Lenc. 2015. MatConvNet Convolutional neural Networks for MATLAB. InProceeding of the ACM Int. Conf. on Multimedia.