ARTIFICIAL NEURAL NETWORKS
WHAT IS ARTIFICIAL NEURAL NETWORK?
👉Artificial Neural Network or neural networks are computational algorithms that intended to simulate the behavior of biological systems composed of neurons.
👉ANNs are computational models inspired by an animal’s central nervous systems. It is capable of machine learning as well as pattern recognition.
👉A neural network is an oriented graph. It consists of nodes which in the biological analogy represent neurons, connected by arcs.
👉It corresponds to dendrites and synapses. Each are associated with a weight at each node. A neural network is a machine learning algorithm based on the model of a human neuron. The Human brain consists of millions of neurons.
👉It is an information processing techniques. It works like the way human brain processes information.
👉ANN includes a large number of connected processing units that work together to process information. They also generate meaningful results from it.
LAYERS IN ARTIFICIAL NEURAL NETWORKS👉
➡️ INPUT LAYER:
🔸The activity of the input units represents the raw information that can feed into the network.
➡️ HIDDEN LAYER:
🔸Hidden layer is used to determine the activity of each hidden unit.
🔸The activities of the input units and weights depend on the connections between he input and the hidden units.
🔸There maybe one or more layers.
➡️ OUTPUT LAYER:
🔸The behavior of the output unit depends on the activity of the hidden units and the weights between the hidden and the output units.
ADVANTAGES OF ARTIFICIAL NEURAL NETWORKS👉
🔸Problems in ANN are represented by attribute value-pairs.
🔸ANNs are used for problems having he target function, output may be discrete-valued, real-valued, or a vector of several real attributes.
🔸ANNs learning methods are quite robust to noise in the training data .The training examples may contain errors, which do not affect the final output.
🔸It is used where the fast evaluation of the learned target function required.
DISADVANTAGES OF ARTIFICIAL NEURAL NETWORKS👉
- Hardware dependence :
♦ Neural Networks require processors with parallel processing power, by their structure.
♦ For this reason, the realization of the equipment is dependent.
2. Unexplained functioning
♦This is the most important problem of ANN
♦When ANN gives a probing solution, it doesn't give a clue as to why and how.
3.The difficulty of showing the problem to the network:
♦ ANNs can work with numerical information
♦ Problems have to be translated into numerical values before being introduced to ANN.
CHARACTERISTICS 👉
♦It is neurally implemented mathematical model.
♦It contains large number of interconnected processing elements called neurons to do all the operations.
♦Information stored in the neurons is basically the weighted linkage of neurons.
♦The input signals arrive at the processing elements through connections and connecting weights.
APPLICATIONS AREAS OF NEURAL NETWORKS👉
Speech recognition
Speech occupies a prominent role in human-human interaction. Therefore, it is natural for people to expect speech interfaces with computers. To ease the communication barrier, a simple solution could be communication in a spoken language that is possible for the machine to understand.
Character and Face recognition:
Face recognition is a visual pattern recognition problem. In detail, a face recognition system with the input of an arbitrary image will search in database to output people’s identification in the input image.Facial components, such as eyes, nose, and mouth and facial outline, are located; based on the location points, the input face image is normalized with respect to geometrical properties, such as size and pose, using geometrical transforms or morphing.
The face is usually further normalized with respect to photometrical properties such illumination and gray scale. After a face is normalized geometrically and photometrically, feature extraction is performed to provide effective information that is useful for distinguishing between faces of different persons and stable with respect to the geometrical and photometrical variations.
It is a problem which falls under the genera area of Pattern recognition many neural networks have been developed for automatic recognition of handwritten characters either letters or digits.
Signature verification application:
Signatures are useful ways to authorize and authenticate a person in legal transactions. Signature verification technology utilizes the distinctive aspects of the signature to verify the identity of individuals. The technology examines the behavioral components of the signature, such as stroke order, speed and pressure, as opposed to comparing visual images of signatures.
It is a non-vision based technique. For this application, the first approach is to extract the feature or rather the geometrical feature set representing the signature. This trained neural network will classify the signature as being genuine or forged under the verification stage.
USE-CASES OF NEURAL NETWORKS
📌Instagram is a social media platform for visual media sharing and it is increasingly being adopted by traditional media platforms as an extension or addition to their activities. In such a scenario the analysis of the popularity or traction of the Instagram posts becomes important for estimation of reach etc. Also, in a commercial scenario when sponsored content is to be published on the medium it is important to be able to coarsely predict the reach of a particular post for price fixation .
📌Instagram, acquired by Facebook back in 2012, uses deep learning by making use of a connection of recurrent neural networks to identify the contextual meaning of an emoji — which has been steadily replacing slangs (for instance, a laughing emoji could replace “rofl”).
By algorithmically identifying the sentiments behind emojis, Instagram creates and auto-suggests emojis and emoji related hashtags. This may seem like a minor application of AI, but being able to interpret and analyze this emoji-to-text translation at a larger scale sets the basis for further analysis on how people use Instagram.
LET’S FOUND OUT HOW INSTRAGRAM ACTUALLY USES NEURAL NETWORK
➡️ Hinton et. al. introduced the concept of deep learning in , which has found successful applications in multiple areas establishing it as a new field of machine learning research. The motivation lies in the establishment neural network which is able to break down complex learning problems into simpler representations.
For the work presented in this paper metrics are first extracted by crawling of the Instagram channel using an application programing interface (API). These include parameters which log the interactions of the media consumers with the posts like ‘comments‘, ‘likes‘, ‘comments‘ etc. The ‘tag‘ metric is a discrete set where the entries although related, might Fig. 1. Generalized block diagram of an Auto-Encoder.
➡️It have different forms. For instance, a post tagged #watches is related to a post tagged #seiko by virtue of ‘Seiko‘ being a manufacturer of watches. However, lexicographically they have little inter-relation. To overcome this a tree of words were constructed which allowed similar tags to be automatically grouped into taxonomic categories with the underlying assumption that if the same tags repeatedly appear together in posts, they might be related.
➡️In the proposed methodology we use a Stacked-Auto Encoder (SAE) network to integrate these metrics collected and learn an optimal model that will allow the prediction of the popularity of the post. A generalized auto-encoder is represented in Figure 1. As shown, it consists of input encoder layers, a central representational layer and output decoder layers. It works by learning an efficient representation of the input data. It is a fully connected neural network who’s outputs are tied to its inputs in the training phase. Thus, after training the networks weights are able to closely approximate the input data, and in the ideal case, learn a perfect representation of it.
➡️A sparse auto-encoder has a large number of hidden units per layer but only a small percentage produce significant activation. Such mechanisms have been shown to have learned more useful representations of the data than the original input. Variations of this technique have been successfully used in different fields to extract better features from data without need for pre-training [10]. Features are extracted after training from the central representational layer. These are then given as input the multi-layer perceptron (MLP) part of the network with sigmoid non-linearities. This computes the predicted quantized performance metric. II. METHODOLOGY Data from the GQ India Instagram account was extracted using the (API) provided by Instagram. The API returns 32 requests per invocation in a JavaScript Object Notation (JSON) format: 65 such JSON objects were aggregated to collect the full corpus of the data from the account consisting of a total of 1280 entries or posts.
➡️Each post is represented by several parameters in the JSON, such as interactions (’comments’ and ’likes’) on post, users who have commented, tags, type of post, filters applied on the post etc. We filtered parameters relevant to the current study from this corpus and used it for training of a deep neural network. The number of likes in the post were granulized to groups of 25 and labeled as classes. For example, posts with number of likes between 0–25 were labeled as Class 1, and so on. A. Feature Selection Feature selection is an integral part of design of a learning algorithm. The following features are extracted from the raw data to represent each post: The type of filter applied, If location data is available for that particular post, The creation time of the post (expressed as a UNIX seconds timestamp), week in the year and day of the week when the post was created, hour of creation, the URL of the image itself, number of users tagged in the post, the textual caption attached with the post, the length of the caption, the number of tags expressed in the caption and a list of the tags itself. This is tabulated in Table I along with a sample data-point of each. Each of these parameters can influence the popularity of the post as a result of human perception and psychology.
➡️Instance, posts made during work hours on weekdays are likely to be less visible as compared to one made on a week-end day evening. Thus parameters related to the creation instance of the post were included as features. Studies have shown that the presence of human faces in the image [ can influence the engagement. We use the number of tagged users as an indirect measure for the presence of faces in the image. A large portion of the posts made by GQ India correspondents use the re-post mechanism, i.e. they copy images from posts created by other users. The filter parameter indicated in data obtained from Instagram only catalogues the last filter applied in their mobile interface and thus cannot be used as a reliable indicator of the mood or visual style of the image. Several sophisticated algorithms exist in literature that allows the extraction of the image style. However, for this study we used a relatively simple approach.
➡️ Each image was downloaded from the provided URL and basic color and brightness information was extracted. The percentage of Red, Blue and Green content in the image was evaluated and expressed on a 0 to 1 scale. The image was then converted to greyscale using a fast, contrast enhancing, colour to grayscale conversion algorithm and the average brightness of the resultant image was expressed between 0 to1 and used as a descriptor of the image.
HOW THE ALGORITHM WORKS👉
Feature selection
Feature selection is an integral part of design of a learning algorithm. Studies have shown that the presence of human faces in the images can influence the engagement. We use the number of tagged users as an indirect measure for the Presence of Human Faces in the Images. We used a relatively simple approach to get at the cause of the popularity of the post. We will use a filter that catalogues the last filter applied in Instagram’s mobile interface.
Network Architecture
We use a deep neural network (DNN) with four hidden layers followed by a multi-layer perceptron (MLP) network. The network comprises rectified linear unit (ReLU) hidden layers to implement the learning algorithm. In practice this has some drawbacks. During back-propogation if the input data to the neuron is always positive, then the gradient weights will either become all positive, or all negative.
Training and Testing
The dataset was randomly split with 70% of the data being used for training and 30% used for testing. A 3 fold cross validation was applied apart from the random shuffling of data on each iteration. For the training phase the base learning rate (α) is selected as 0.001 and it is stepped every 600 iterations.
RESULTS
The network was able to achieve an accuracy of 88% (on average) converging to a stable state within 1500 iterations. With a granularity of 25 likes per class, this performance is acceptable for commercial applications such as prediction of popularity of a sponsored post, hence price fixation. The performance of the algorithm can be further improved by improvement in the feature collection. Instead of relying on the tagged users to detect the presence or absence of faces, a face detection algorithm can be used to populate that data.
So here in this blog we have discussed that how artificial neural network works and also how Instagram uses artificial neural network as a use-case.
LIKE SHARE AND COMMENT!
HAPPY LEARNING😃
BY -SHRISTI SHARMA:)
