365 Data Science

An Atificial Intelligence developed its own non-human language

The buried line in the new Facebook report about one-on-one chatbots conversations provides a great glimpse into the future of the language.

In the report, researchers at the Facebook Artificial Intelligence Research Lab describe the use of machine learning to negotiate with their “dialogue agents.” (And bots seem to be very good at dealing with it.)

To know about: Six ways AI is revolutionizing e-commerce

At one point, the researchers wrote, they had to adjust one of their models, otherwise the bot-to-bot dialogue “developed their language for negotiation, which led to the disagreement from human language to agents. “Instead they had to use the so-called fixed supervised model.

In other words, the model that allowed the two bots to communicate — and to use machine learning to continually redirect strategies for that conversation — led the bots to communicate in their non-human language. If it doesn’t surprise you and wonders about the future of machines and humanity.

Why the New Language?

AI agents work on a “reward principle” such as positive reinforcement. They are expected to take some sort of advantage in completing tasks and actions. Speaking English gave them no advantage, so they devised a new language to make communication more effective.

To know about: Why AI Is a Practical Solution for Pharma?

This language can be compared to the use of shorthand for humans. This is the way to create a more efficient and less time-consuming solution for communicating their ideas. However, we do not have that luxury for AI, as we humans have to teach other humans about shorthand

To Know about:5G Driving the Evolution of AI

To be clear, Facebook’s chatty bots are not evidence of the arrival of the singular. Not even close. They demonstrate how machines have redefined public perception of many kingdoms that once believed to be a particularly human-like language.

Already, there is a good deal of work in machine learning research, which often involves accumulating massive data on the neural net, and then examining the output to understand how the machine thinks. The point that machines can compose their non-human ways of communicating is a surprising reminder of how little we know even when designing these systems.

Trending AI Articles:

1. AI for CFD: Intro (part 1)

2. Using Artificial Intelligence to detect COVID-19

3. Real vs Fake Tweet Detection using a BERT Transformer Model in few lines of code

4. Machine Learning System Design

Facebook researchers wrote in their paper, “In particular in exploring other logical strategies and improving the diversity of pronunciations without distinguishing them from the human language.”

What’s the big deal?

AI serviced agents who speak in their language pose a threat to humans because we do not know what is being said in their conversion. This leads to a loss of control over artificial intelligence.

It is important to continue to make advances in science and technology, but on the other hand, if we lose control, we can at any moment have the classic case of Robot.There is still insufficient evidence to determine whether this new language poses a threat to agents that override their operators, but it may make development more difficult.

Don’t forget to give us your ? !

---

AN ARTIFICIAL INTELLIGENCE DEVELOPED ITS OWN NON-HUMAN LANGUAGE was originally published in Becoming Human: Artificial Intelligence Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.

Via https://becominghuman.ai/an-artificial-intelligence-developed-its-own-non-human-language-cb3ddfecffda?source=rss—-5e5bef33608a—4

In this blog, we are going to look at RNN ie. Recurrent Neural Network. I’ll be sharing the theory and then we’ll solve a real-time problem using RNN.

RNN is used for sequential data such as Time series data, Heartbeat data.

About RNN

If anyone asks you 1,2,3,4,5, ??? You can easily tell the next number. But RNN needs to know the previous history of outputs. One easy way to do this is to simply feed it’s output back into itself as an output.

Cells that are functions of inputs from previous time steps are also known as memory cells. Rnn is also flexible in their inputs and outputs, for both sequences and single vector values.

Vanishing Gradients

For complex data, we need deep layer. The issue can arise during backpropagation. Backpropagation goes backwards from output to the input layer, propagating the error gradient. For deeper networks issues can arise from backpropagation, vanishing and exploding gradients. As you go back to the lower layers gradients often get smaller, eventually causing weights to never change at lower levels. The opposite can also occur, gradients explode on the way back, causing issues.

Now let’s start some hands-on with RNN.

Trending AI Articles:

1. AI for CFD: Intro (part 1)

2. Using Artificial Intelligence to detect COVID-19

3. Real vs Fake Tweet Detection using a BERT Transformer Model in few lines of code

4. Machine Learning System Design

About Dataset

Release: Advance Monthly Sales for Retail and Food Services
Units: Millions of Dollars, Not Seasonally Adjusted

Frequency: Monthly

The value for the most recent month is an advance estimate that is based on data from a subsample of firms from the larger Monthly Retail Trade Survey. The advance estimate will be superseded in the following months by revised estimates derived from the larger Monthly Retail Trade Survey. The associated series from the Monthly Retail Trade Survey is available at https://fred.stlouisfed.org/series/MRTSSM448USN

Information about the Advance Monthly Retail Sales Survey can be found on the Census website at https://www.census.gov/retail/marts/about_the_surveys.html

Suggested Citation: U.S. Census Bureau, Advance Retail Sales: Clothing and Clothing Accessory Stores [RSCCASN], retrieved from FRED, Federal Reserve Bank of St. Louis; https://fred.stlouisfed.org/series/RSCCASN, November 16, 2019.

Advance Retail Sales: Clothing and Clothing Accessory Stores

Now let’s jump into coding.

Import all the necessary libraries.

import pandas as pd

import numpy as np

import matplotlib.pyplot as plt

from sklearn.preprocessing import MinMaxScaler

from tensorflow.keras.preprocessing.sequence import TimeseriesGenerator

from tensorflow.keras.models import Sequential

from tensorflow.keras.layers import Dense, LSTM

from tensorflow.keras.callbacks import EarlyStopping

Loading the dataset

dataset = pd.read_csv(‘RSCCASN.csv’, index_col=’DATE’, parse_dates=True)

Checking info of the dataset

dataset.info()

Checking description of the data

Checking first 5 columns of the dataset

dataset.head()

Now let’s split the dataset into training and testing

test_sp=len(dataset)-18

train = dataset.iloc[:test_sp]

test = dataset.iloc[test_sp:]

Apply MinMaxScaler to fit and transform the training and testing data

scaler = MinMaxScaler()

scaler.fit(train)

scaled_train = scaler.transform(train)

scaled_test = scaler.transform(test)

Apply TimeSeriesGenerator to predict the sales beyond the dataset date and time.

length = 12

generator = TimeseriesGenerator(scaled_train,scaled_train,length=length, batch_size=1)

Create LSTM model

model = Sequential()

model.add(LSTM(100, activation=’relu’, input_shape=(length, 1)))

model.add(Dense(1))

model.compile(optimizer=’adam’, loss=’mse’)

Summary of our model

model.summary()

Apply EarlyStoping to prevent the model from overfitting

early_stop = EarlyStopping(monitor=’val_loss’,patience=2)

validation_generator = TimeseriesGenerator(scaled_test,scaled_test,length=length, batch_size=1)

model.fit_generator(generator,epochs=10,

validation_data=validation_generator,

callbacks=[early_stop])

Now let’s predict the output from our model and tally with the data value

n_features = 1

test_predictions = []

first_eval_batch = scaled_train[-length:]

current_batch = first_eval_batch.reshape((1, length, n_features))

for i in range(len(test)):

# get prediction 1 time stamp ahead ([0] is for grabbing just the number instead of [array])

current_pred = model.predict(current_batch)[0]

# store prediction

test_predictions.append(current_pred)

# update batch to now include prediction and drop first value

current_batch = np.append(current_batch[:,1:,:],[[current_pred]],axis=1)

prediction = scaler.inverse_transform(test_predictions)

test['Predictions'] = prediction

test

In the above screenshot, you can view the predicted value and dataset value together.

For complete source code, you can access my Github repository here.

I hope you like this blog. Feel free to share your thoughts in the comment section and you can also connect with me in:-
Linkedin — https://www.linkedin.com/in/shreyak007/
Github — https://github.com/Shreyakkk
Twitter — https://twitter.com/Shreyakkkk

Happy Learning.

Thank You.

Don’t forget to give us your ? !

---

A to Z About Recurrent Neural Network (RNN). was originally published in Becoming Human: Artificial Intelligence Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.

Via https://becominghuman.ai/a-to-z-about-recurrent-neural-network-rnn-43562f7c49c4?source=rss—-5e5bef33608a—4

This is a slightly more intricate example of MCMC, compared to many with a fairly simple model, a single predictor (maybe two), and not much else, which highlights a couple of issues and tricks worth noting for a handwritten implementation.

Originally from KDnuggets https://ift.tt/2ZfIEUz

Introduction

Computer vision is a field under development that is improving quickly with new researches and ideas. Millions have been invested in technology and the benefits has spread in many fields like autonomous driving, health, security and banking. One of the works that meant a huge leap towards the improvements of the networks that processes images and makes computer vision posible was the winner from the ILSVRC 2012 competition.

The work of Krizhevsky, Sutskever and Hinton in his paper ImageNet Classification with Deep Convolutional Neural Networks is one of the most influential in computer vision. They called the architecture of the network AlexNet.

The purpose of this writing is to explain AlexNet, what it brings over previous convolutional network architectures and what procedures were involved to be so influential in the field.

Trending AI Articles:

1. AI for CFD: Intro (part 1)

2. Using Artificial Intelligence to detect COVID-19

3. Real vs Fake Tweet Detection using a BERT Transformer Model in few lines of code

4. Machine Learning System Design

Before continuing, is important to know that in the field of computer vision, the convolutional neural networks are of great importance. The inclusion of convolutional layers in a network allows to process a huge amount of parameters and perform a classification by highlighting the most important features of an image. An architecture of a network may differ in it’s resulting test error in the task of classifying an images under a category and also how good is its performance over datasets which has not be included in the training.

Procedures

What did they do to be winners of the competition and have such good results?

• First it is important to know which data they used. In the case of the ILSVRC competition, it has a subset of the ImageNet with roughly 1000 images for each of the 1000 categories. They took advantage of the ILSVRC-2010 data set which has test set labels to perform the majority of their experiments. The pictures from ImageNet are of variable resolution but their system required constant dimensions so they cropped them to 256 x 256 each with RGB.
• They put several novels on their architecture that today are very used. The use of the ReLu nonlinearity as an activation for the neurons is a very important feature that allowed faster training over the use of Tanh or sigmoid.
• Two GTX 580 GPUs with 3GB of memory with parallelization scheme helped them to train the over 1.2 million examples they used for training the networks.

• ReLu makes unnecessary the normalization of input to prevent saturating but they used Local Response Normalization (LRN) that helps normalize data. That reduced top-1 and top-5 error rates by 1.4% and 1.2%, respectively.
• To avoid overfitting they used data augmentation, dropout and weight decay (0.0005) as regularization methods. To optimize, they separated the process in batch sizes of 128 examples using gradient descent, also used momentum of 0.9.

The architecture

The network contains 8 layers in total, 5 are convolutional layers and 3 are fully connected layers. At the end of each layer ReLu activation is performed except for the last one which outputs with a softmax with a distribution over the 1000 class labels. Dropout is applied in the first two fully connected layers. As the figure from above shows, they also apply Max-pooling after the first, second and fifth convolutional layer. Remember that each input image has dimensions of 224 x 224 x 3 since are color images although later discussion of the paper says that 227 x 227 x 3 makes more sense for the dimension calculations for the next of the network.

Results

• I think that what is most important from their work is the use of ReLu which will mark a general use of this activation function after the paper for most works of convolutional neural networks. They got from testing with CIFAR-10 dataset a 25% training error rate that was six times faster than Tanh.
• I was also very important for this field the use of multiple GPUs since training requieres a lot of iteration and time. Faster computation means more testing and that way more discovers.
• The results in numbers they obtained was of 37.5% for top-1 and 17.0% for top-5 error rate. For ImageNet, these two errors are report, top-1 is the normal error rate for classification and top-5 is the fraction of test images which correct label is not among the five labels considered most probable my the model. This result surpassed results of previous convolutional neural networks.

Conclusion

AlexNet is a work of supervised learning and got very good results. It is not easy to have low classification errors without having of overfitting. They say that removing one convolutional layer from their network would reduce drastically the performance so its no easy task to choose the architecture. It was also important the selection of methods like dropout and data augmentation that helped the performance of the network. Finally, the tools they used sped up a training process which otherwise would be very daunting with 1.2 million high resolution images.

Personal Notes

• What impresses a lot is the carefully crafted result of their network, it involved a lot of testing and decision because every run of the model would take a lot of time. The idea of put one element and a layer where it is made a difference and every % reduction of error counted.
• It is no easy work to avoid overfitting when you build a bigger neural network while you keep a low classification error. Regularization is very important.
• There are many opensource datasets with labeled data like CIFAR and ImageNet that didn’t exists that today makes possible to train models and facilitate investigations.
• Hackatons and challenges encourages the generation of ideas. They bring solutions to common problems on society so its nice that such competitions are promoted and more people participate.
• AlexNet is still relevant today but it is true that there are new researches. It is important for someone who wants to dig into Machine Learning field to know how to read papers and gather the information on how the networks depicted were constructed.

References

• ImageNet Classification with Deep Convolutional Neural Networks
• ImageNet Large Scale Visual Recognition Competition 2012 (ILSVRC2012)

Don’t forget to give us your ? !

---

Explanation of AlexNet and its leap for CNNs was originally published in Becoming Human: Artificial Intelligence Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.

Via https://becominghuman.ai/explanation-of-alexnet-and-its-leap-for-cnns-910244cb4a90?source=rss—-5e5bef33608a—4

What is big tech and the government doing about deepfakes? We look at where the technology is in development, what regulations are in place, and how it could impact 2020.

Deepfakes are synthetic media, usually videos, created with deep learning technology. By manipulating images, videos, and voices of real people, a deepfake can portray someone doing things they never did, or saying things they never said.

By feeding a machine learning model thousands of target images, a deepfake algorithm can learn the details of a person’s face. With enough training data, the algorithm can then predict what that person’s face would look like when mimicking the expressions of someone else. A similar process is used for training deepfake algorithms to mimic the accent, intonation, and tone of a person’s voice.

The Public Response to Deepfakes

The start of 2020 came with an interesting shift in response to deepfake technology, when Facebook announced a ban on manipulated videos and images on their platforms. Facebook said it would remove AI-edited content that was likely to mislead people, but added that the ban doesn’t include parody or satire. Lawmakers, however, are skeptical as to whether the ban goes far enough to address the root problem: the ongoing spread of disinformation.

The speed and ease with which a deepfake can be made and deployed, as shown in this article by Ars Technica, have many worried about misuse in the near future, especially with an election on the horizon for the U.S. Many in America, including military leaders, have also weighed in with worries about the speed and ease with which the tech can be used. These concerns are heightened by the knowledge that deepfake technology is improving and becoming more accessible.

This news is the latest in a series of initiatives to detect and regulate deepfake releases. Exactly how to handle them is an ongoing discussion. Twitter announced in November that it intended to draft a deepfakes policy. In December, Facebook announced The Deepfake Detection Challenge in conjunction with tech giants like Microsoft and Amazon. The challenge offers financial rewards for the building of technology that helps to detect manipulated media. Google, too, has contributed a dataset of deepfakes for the purposes of developing better technology to detect them.

The Year of Consumerized Deepfakes?

This growing concern around deepfake detection has not stopped its move into social media. In fact, recent technological developments on social media platforms have people talking about the idea of consumerized deepfakes. Code found in mobile apps Douyin and TikTok has revealed technology allowing users to insert their face into videos starring other people. Though the application is still unreleased, it’s a prime example of how social media platforms are utilizing deepfake technology.

Snap, the company behind Snapchat, has also reportedly acquired AI Factory, an image and video recognition start-up. Reports state that Snap used AI Factory’s technology for a new face swapping feature, raising some concerns around the possibility of deepfake usage. These applications of deepfake technology will likely continue as social media apps look to entice new users.

Trending AI Articles:

1. AI for CFD: Intro (part 1)

2. Using Artificial Intelligence to detect COVID-19

3. Real vs Fake Tweet Detection using a BERT Transformer Model in few lines of code

4. Machine Learning System Design

Improved Detection

However, together with worries surrounding the misuse of this technology, deepfake detection is also improving. Teams at Microsoft Research and Peking University recently proposed Face X-Ray, a tool that recognizes whether a headshot is fake. The tool detects points in a headshot where one image has been blended with another to create a forgery. Though the technology is a step in the right direction, the researchers also state their technology is unable to detect a wholly synthetic image. This would make it weak against adversarial samples.

However, most researchers are pointing to improved education as the best mode of defense. They recommend keeping in mind the following questions when viewing video content:

• Is the video bizarre or exceptional?
• Is the video quality low or grainy?
• How short is the video? (30 to 60 seconds long is common for current deepfakes.)
• Is the content visually or aurally strange? (blurry faces, strange lighting, voices/lips out of sync, etc.)

What’s Next?

The above news articles point to a clear trend: deepfake technology isn’t going away. Its use in the future will range from benign and novel to potentially destructive and damaging. Development and detection is quickly becoming a race as detection methods struggle to keep up with improving technology. The result? Regular consumers of news media will find it more difficult to differentiate the real news from the fake. This makes a strong case for improved education, which means focusing on how to educate, and where to start.

With this in mind, how tech companies like Twitter and Facebook regulate deepfake technology will have a huge impact on how people consume and understand it. 2020 will likely bring more discussion around regulation, and more development for entertainment and social media. This will go hand in hand with improved technology for both implementation and detection. But for a deeper look at the potential threats and countermeasures to deepfake technology, see our comprehensive article here.

Along with generative neural networks and synthetic voice technology, Deepfakes are one of many machine learning trends we expect to see in the news in 2020. Read our report on the state of facial recognition bias here.

Original article reposted with permission.

Don’t forget to give us your ? !

---

A Look at Deepfakes in 2020 was originally published in Becoming Human: Artificial Intelligence Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.

Via https://becominghuman.ai/a-look-at-deepfakes-in-2020-13d3fe2b6ef7?source=rss—-5e5bef33608a—4

There exist so many great computational tools available for Data Scientists to perform their work. However, mathematical skills are still essential in data science and machine learning because these tools will only be black-boxes for which you will not be able to ask core analytical questions without a theoretical foundation.

Originally from KDnuggets https://ift.tt/2NviYh0

An article highlighting things I’ve learned in the real world about Data science.

Originally from KDnuggets https://ift.tt/31jLXwH

Watch the evolution of the top 10 most popular data science tools based on KDnuggets software polls from 2000 to 2019.

Originally from KDnuggets https://ift.tt/37ZRNEJ

Check out this article for a brief summary on what LynxKite is, where it is coming from and how it can help with your data science projects.

Originally from KDnuggets https://ift.tt/2VjJ0YX

DeepMind’s research shows how to understand the role of individual neurons in a neural network.

Originally from KDnuggets https://ift.tt/38bReYF